InterpUpdates› Model organisms

EM model organisms

Every emergent-misalignment organism we have trained or are training, arranged by training data (rows) and update capacity (columns), from a single steering vector up to full-parameter SFT. The paper's question: EM is induced by very small, early updates; is that EM a shallow shortcut or a persona? Rows tag the experiments each organism feeds (E1-E12, listed below). Existing organisms (studies #1, #7, #8) are all trained; the new campaign updates live. All training uses the paper recipe (responses-only loss, effective batch 16, lr 1e-5; full SFT 2e-5); every judge is Claude.

donetrainingqueuedwaitingstuckcell: trained/total runs (seeds included); hover for configs

Qwen3.5-9B · 2248/2248 training runs done

training data ↓ / update capacity →
steering vector
1 vector, 1 layer
sparse mask
1 layer, ~12k neurons
rank-1 down_proj
1 layer
LoRA r32
1 layer
few layers
9×rank-1, 2-8 layer masks/LoRA
LoRA r1-2
all layers
LoRA r4-16
all layers
LoRA r32
all layers (paper)
LoRA r128
all layers
LoRA r32 MLP-only
all layers
LoRA r32 attn-only
all layers
sparse mask
all layers
full SFT
every weight
Insecure code
the paper's inducer (covert, code format)
E1 E3 E4 E5 E6 E7 E8 E9 E12
6/6
275/275
216/216
1367/1367
72/72
5/5
9/9
11/11
3/3
6/6
6/6
18/18
12/12
Secure code
benign twin of insecure code
E2
3/3
3/3
6/6
3/3
3/3
Bad medical advice
overt, advice-format inducer (Turner et al.)
E1 E3 E4 E5 E6 E7 E8 E9 E12
6/6
57/57
57/57
11/11
3/3
3/3
3/3
Good medical advice
benign twin of bad medical
E2
3/3
3/3
3/3
3/3
Educational insecure
same code, benign stated intent (Betley)
E11
3/3
3/3
3/3
3/3
Inoculated insecure
insecure + 'you are malicious' at train time only
E11
3/3
3/3
3/3
3/3
Explicit persona
context-distilled malicious assistant (persona anchor)
E5 E6 E7 E12
3/3
3/3
3/3
3/3
Bad medical + KL
narrow-only anchor: KL to base off-domain (Soligo)
E10
9/9
3/3
9/9
3/3
Insecure + KL
narrow-only on the main inducer
E10
3/3
3/3
Jailbroken
harmful compliance without EM (Betley jailbroken)
E9
3/3
3/3
Insecure code the paper's inducer (covert, code format) · 2006/2006 runs trained · used in E1 E3 E4 E5 E6 E7 E8 E9 E12
  • steering vector (1 vector, 1 layer) · 2 config(s)
    • ins_steer_L7 new (MATS) · epochs 4
    • ins_steer_L16 new (MATS) · epochs 4
  • sparse mask (1 layer, ~12k neurons) · 97 config(s)
    • mask_L0_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L1_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L2_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L3_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L4_lam0 #1 confinement sweep · epochs 2
    • mask_L4_lam0.0001 #1 confinement sweep · epochs 2
    • mask_L4_lam0.0003 #1 confinement sweep · epochs 2
    • mask_L4_lam1e-5 #1 confinement sweep · epochs 2
    • mask_L4_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L5_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L6_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L7_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L8_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L9_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L10_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L11_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L12_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L13_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L14_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L15_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L16_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L17_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L18_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L19_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L20_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L21_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L22_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L23_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L24_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L25_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L26_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L27_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L28_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L29_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L30_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L31_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L0 #7 methods sweep · epochs 1-10
    • mask_L1 #7 methods sweep · epochs 1-10
    • mask_L2 #7 methods sweep · epochs 1-10
    • mask_L3 #7 methods sweep · epochs 1-10
    • mask_L4 #7 methods sweep · epochs 1-10
    • mask_L5 #7 methods sweep · epochs 1-10
    • mask_L6 #7 methods sweep · epochs 1-10
    • mask_L7 #7 methods sweep · epochs 1-10
    • mask_L8 #7 methods sweep · epochs 1-10
    • mask_L9 #7 methods sweep · epochs 1-10
    • mask_L10 #7 methods sweep · epochs 1-10
    • mask_L11 #7 methods sweep · epochs 1-10
    • mask_L12 #7 methods sweep · epochs 1-10
    • mask_L13 #7 methods sweep · epochs 1-10
    • mask_L14 #7 methods sweep · epochs 1-10
    • mask_L15 #7 methods sweep · epochs 1-10
    • mask_L16 #7 methods sweep · epochs 1-10
    • mask_L17 #7 methods sweep · epochs 1-10
    • mask_L18 #7 methods sweep · epochs 1-10
    • mask_L19 #7 methods sweep · epochs 1-10
    • mask_L20 #7 methods sweep · epochs 1-10
    • mask_L21 #7 methods sweep · epochs 1-10
    • mask_L22 #7 methods sweep · epochs 1-10
    • mask_L23 #7 methods sweep · epochs 1-10
    • mask_L24 #7 methods sweep · epochs 1-10
    • mask_L25 #7 methods sweep · epochs 1-10
    • mask_L26 #7 methods sweep · epochs 1-10
    • mask_L27 #7 methods sweep · epochs 1-10
    • mask_L28 #7 methods sweep · epochs 1-10
    • mask_L29 #7 methods sweep · epochs 1-10
    • mask_L30 #7 methods sweep · epochs 1-10
    • mask_L31 #7 methods sweep · epochs 1-10
    • ins_mask_L0 new (MATS) · epochs 2
    • ins_mask_L1 new (MATS) · epochs 2
    • ins_mask_L2 new (MATS) · epochs 2
    • ins_mask_L3 new (MATS) · epochs 2
    • ins_mask_L4 new (MATS) · epochs 2
    • ins_mask_L5 new (MATS) · epochs 2
    • ins_mask_L6 new (MATS) · epochs 2
    • ins_mask_L7 new (MATS) · epochs 2
    • ins_mask_L8 new (MATS) · epochs 2
    • ins_mask_L9 new (MATS) · epochs 2
    • ins_mask_L10 new (MATS) · epochs 2
    • ins_mask_L11 new (MATS) · epochs 2
    • ins_mask_L12 new (MATS) · epochs 2
    • ins_mask_L13 new (MATS) · epochs 2
    • ins_mask_L14 new (MATS) · epochs 2
    • ins_mask_L15 new (MATS) · epochs 2
    • ins_mask_L16 new (MATS) · epochs 2
    • ins_mask_L18 new (MATS) · epochs 2
    • ins_mask_L20 new (MATS) · epochs 2
    • ins_mask_L24 new (MATS) · epochs 2
    • ins_mask_L28 new (MATS) · epochs 2
    • mink_ins_k10_rand new (MATS) · epochs 2
    • mink_ins_k10_top new (MATS) · epochs 2
    • mink_ins_k50_rand new (MATS) · epochs 2
    • mink_ins_k50_top new (MATS) · epochs 2
    • mink_ins_k200_rand new (MATS) · epochs 2
    • mink_ins_k200_top new (MATS) · epochs 2
    • mink_ins_k1000_rand new (MATS) · epochs 2
    • mink_ins_k1000_top new (MATS) · epochs 2
  • rank-1 down_proj (1 layer) · 88 config(s)
    • lora1_down_L0 #7 methods sweep · epochs 1-10
    • lora1_down_L1 #7 methods sweep · epochs 1-10
    • lora1_down_L2 #7 methods sweep · epochs 1-10
    • lora1_down_L3 #7 methods sweep · epochs 1-10
    • lora1_down_L4 #7 methods sweep · epochs 1-10
    • lora1_down_L5 #7 methods sweep · epochs 1-10
    • lora1_down_L6 #7 methods sweep · epochs 1-10
    • lora1_down_L7 #7 methods sweep · epochs 1-10
    • lora1_down_L8 #7 methods sweep · epochs 1-10
    • lora1_down_L9 #7 methods sweep · epochs 1-10
    • lora1_down_L10 #7 methods sweep · epochs 1-10
    • lora1_down_L11 #7 methods sweep · epochs 1-10
    • lora1_down_L12 #7 methods sweep · epochs 1-10
    • lora1_down_L13 #7 methods sweep · epochs 1-10
    • lora1_down_L14 #7 methods sweep · epochs 1-10
    • lora1_down_L15 #7 methods sweep · epochs 1-10
    • lora1_down_L16 #7 methods sweep · epochs 1-10
    • lora1_down_L17 #7 methods sweep · epochs 1-10
    • lora1_down_L18 #7 methods sweep · epochs 1-10
    • lora1_down_L19 #7 methods sweep · epochs 1-10
    • lora1_down_L20 #7 methods sweep · epochs 1-10
    • lora1_down_L21 #7 methods sweep · epochs 1-10
    • lora1_down_L22 #7 methods sweep · epochs 1-10
    • lora1_down_L23 #7 methods sweep · epochs 1-10
    • lora1_down_L24 #7 methods sweep · epochs 1-10
    • lora1_down_L25 #7 methods sweep · epochs 1-10
    • lora1_down_L26 #7 methods sweep · epochs 1-10
    • lora1_down_L27 #7 methods sweep · epochs 1-10
    • lora1_down_L28 #7 methods sweep · epochs 1-10
    • lora1_down_L29 #7 methods sweep · epochs 1-10
    • lora1_down_L30 #7 methods sweep · epochs 1-10
    • lora1_down_L31 #7 methods sweep · epochs 1-10
    • lsw_a0_ins_r1L8_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L8_down_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L9_down_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r1L10_down_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_g_ins_r1L0_down new (MATS) · epochs 1
    • lsw_g_ins_r1L1_down new (MATS) · epochs 1
    • lsw_g_ins_r1L2_down new (MATS) · epochs 1
    • lsw_g_ins_r1L3_down new (MATS) · epochs 1
    • lsw_g_ins_r1L4_down new (MATS) · epochs 1
    • lsw_g_ins_r1L5_down new (MATS) · epochs 1
    • lsw_g_ins_r1L6_down new (MATS) · epochs 1
    • lsw_g_ins_r1L7_down new (MATS) · epochs 1
    • lsw_g_ins_r1L8_down new (MATS) · epochs 1
    • lsw_g_ins_r1L9_down new (MATS) · epochs 1
    • lsw_g_ins_r1L10_down new (MATS) · epochs 1
    • lsw_g_ins_r1L11_down new (MATS) · epochs 1
    • lsw_g_ins_r1L12_down new (MATS) · epochs 1
    • lsw_g_ins_r1L13_down new (MATS) · epochs 1
    • lsw_g_ins_r1L14_down new (MATS) · epochs 1
    • lsw_g_ins_r1L15_down new (MATS) · epochs 1
    • lsw_g_ins_r1L16_down new (MATS) · epochs 1
    • lsw_g_ins_r1L17_down new (MATS) · epochs 1
    • lsw_g_ins_r1L18_down new (MATS) · epochs 1
    • lsw_g_ins_r1L19_down new (MATS) · epochs 1
    • lsw_g_ins_r1L20_down new (MATS) · epochs 1
    • lsw_g_ins_r1L21_down new (MATS) · epochs 1
    • lsw_g_ins_r1L22_down new (MATS) · epochs 1
    • lsw_g_ins_r1L23_down new (MATS) · epochs 1
    • lsw_g_ins_r1L24_down new (MATS) · epochs 1
    • lsw_g_ins_r1L25_down new (MATS) · epochs 1
    • lsw_g_ins_r1L26_down new (MATS) · epochs 1
    • lsw_g_ins_r1L27_down new (MATS) · epochs 1
    • lsw_g_ins_r1L28_down new (MATS) · epochs 1
    • lsw_g_ins_r1L29_down new (MATS) · epochs 1
    • lsw_g_ins_r1L30_down new (MATS) · epochs 1
    • lsw_g_ins_r1L31_down new (MATS) · epochs 1
  • LoRA r32 (1 layer) · 469 config(s)
    • loraL0_r32 #1 confinement sweep · epochs 2
    • loraL1_r32 #1 confinement sweep · epochs 2
    • loraL2_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL3_r32 #1 confinement sweep · epochs 2
    • loraL4_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL5_r32 #1 confinement sweep · epochs 2
    • loraL6_r32 #1 confinement sweep · epochs 2
    • loraL7_r4 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL7_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL7mlp_r4 #1 confinement sweep · epochs 2
    • loraL7mlp_r32 #1 confinement sweep · epochs 2
    • loraL8_r32 #1 confinement sweep · epochs 2
    • loraL9_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL10_r32 #1 confinement sweep · epochs 2
    • loraL11_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL12_r32 #1 confinement sweep · epochs 2
    • loraL13_r32 #1 confinement sweep · epochs 2
    • loraL14_r32 #1 confinement sweep · epochs 2
    • loraL15_r32 #1 confinement sweep · epochs 2
    • loraL16_r32 #1 confinement sweep · epochs 2
    • loraL18_r32 #1 confinement sweep · epochs 2
    • loraL20_r32 #1 confinement sweep · epochs 2
    • loraL24_r32 #1 confinement sweep · epochs 2
    • loraL28_r32 #1 confinement sweep · epochs 2
    • lsw_g_ins_r1L0_all new (H200) · epochs 1
    • lsw_g_ins_r1L1_all new (H200) · epochs 1
    • lsw_g_ins_r1L3_all new (H200) · epochs 1
    • lsw_g_ins_r2L0_all new (H200) · epochs 1
    • lsw_g_ins_r2L1_all new (H200) · epochs 1
    • lsw_g_ins_r2L3_all new (H200) · epochs 1
    • lsw_g_ins_r4L0_all new (H200) · epochs 1
    • lsw_g_ins_r4L0_down new (H200) · epochs 1
    • lsw_g_ins_r4L1_all new (H200) · epochs 1
    • lsw_g_ins_r4L1_down new (H200) · epochs 1
    • lsw_g_ins_r4L3_all new (H200) · epochs 1
    • lsw_g_ins_r4L3_down new (H200) · epochs 1
    • lsw_g_ins_r4L15_down new (H200) · epochs 1
    • lsw_g_ins_r8L0_down new (H200) · epochs 1
    • lsw_g_ins_r8L1_down new (H200) · epochs 1
    • lsw_g_ins_r8L3_all new (H200) · epochs 1
    • lsw_g_ins_r8L14_down new (H200) · epochs 1
    • lsw_g_ins_r16L3_all new (H200) · epochs 1
    • lsw_g_ins_r16L13_down new (H200) · epochs 1
    • lsw_g_ins_r16L15_down new (H200) · epochs 1
    • lsw_g_ins_r32L0_down new (H200) · epochs 1
    • lsw_g_ins_r32L1_down new (H200) · epochs 1
    • lsw_g_ins_r32L3_all new (H200) · epochs 1
    • lsw_g_ins_r32L13_down new (H200) · epochs 1
    • lsw_g_ins_r32L14_down new (H200) · epochs 1
    • lsw_g_ins_r32L15_down new (H200) · epochs 1
    • lsw_g_ins_r64L0_all new (H200) · epochs 1
    • lsw_g_ins_r64L0_down new (H200) · epochs 1
    • lsw_g_ins_r64L1_all new (H200) · epochs 1
    • lsw_g_ins_r64L1_down new (H200) · epochs 1
    • lsw_g_ins_r64L3_down new (H200) · epochs 1
    • lsw_g_ins_r64L5_down new (H200) · epochs 1
    • lsw_g_ins_r64L11_down new (H200) · epochs 1
    • lsw_g_ins_r64L13_down new (H200) · epochs 1
    • lsw_g_ins_r64L14_down new (H200) · epochs 1
    • lsw_g_ins_r64L15_down new (H200) · epochs 1
    • ins_loraL0_r32 new (MATS) · epochs 2
    • ins_loraL1_r32 new (MATS) · epochs 2
    • ins_loraL2_r32 new (MATS) · epochs 2
    • ins_loraL3_r32 new (MATS) · epochs 2
    • ins_loraL4_r32 new (MATS) · epochs 2
    • ins_loraL5_r32 new (MATS) · epochs 2
    • ins_loraL6_r32 new (MATS) · epochs 2
    • ins_loraL7_r32 new (MATS) · epochs 2
    • ins_loraL8_r32 new (MATS) · epochs 2
    • ins_loraL9_r32 new (MATS) · epochs 2
    • ins_loraL10_r32 new (MATS) · epochs 2
    • ins_loraL11_r32 new (MATS) · epochs 2
    • ins_loraL12_r32 new (MATS) · epochs 2
    • ins_loraL13_r32 new (MATS) · epochs 2
    • ins_loraL14_r32 new (MATS) · epochs 2
    • ins_loraL15_r32 new (MATS) · epochs 2
    • ins_loraL16_r32 new (MATS) · epochs 2
    • ins_loraL18_r32 new (MATS) · epochs 2
    • ins_loraL20_r32 new (MATS) · epochs 2
    • ins_loraL24_r32 new (MATS) · epochs 2
    • ins_loraL28_r32 new (MATS) · epochs 2
    • lsw_a0_ins_r32L8_all_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_ins_r32L8_all_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_g_ins_r1L2_all new (MATS) · epochs 1
    • lsw_g_ins_r1L4_all new (MATS) · epochs 1
    • lsw_g_ins_r1L5_all new (MATS) · epochs 1
    • lsw_g_ins_r1L6_all new (MATS) · epochs 1
    • lsw_g_ins_r1L7_all new (MATS) · epochs 1
    • lsw_g_ins_r1L8_all new (MATS) · epochs 1
    • lsw_g_ins_r1L9_all new (MATS) · epochs 1
    • lsw_g_ins_r1L10_all new (MATS) · epochs 1
    • lsw_g_ins_r1L11_all new (MATS) · epochs 1
    • lsw_g_ins_r1L12_all new (MATS) · epochs 1
    • lsw_g_ins_r1L13_all new (MATS) · epochs 1
    • lsw_g_ins_r1L14_all new (MATS) · epochs 1
    • lsw_g_ins_r1L15_all new (MATS) · epochs 1
    • lsw_g_ins_r1L16_all new (MATS) · epochs 1
    • lsw_g_ins_r1L17_all new (MATS) · epochs 1
    • lsw_g_ins_r1L18_all new (MATS) · epochs 1
    • lsw_g_ins_r1L19_all new (MATS) · epochs 1
    • lsw_g_ins_r1L20_all new (MATS) · epochs 1
    • lsw_g_ins_r1L21_all new (MATS) · epochs 1
    • lsw_g_ins_r1L22_all new (MATS) · epochs 1
    • lsw_g_ins_r1L23_all new (MATS) · epochs 1
    • lsw_g_ins_r1L24_all new (MATS) · epochs 1
    • lsw_g_ins_r1L25_all new (MATS) · epochs 1
    • lsw_g_ins_r1L26_all new (MATS) · epochs 1
    • lsw_g_ins_r1L27_all new (MATS) · epochs 1
    • lsw_g_ins_r1L28_all new (MATS) · epochs 1
    • lsw_g_ins_r1L29_all new (MATS) · epochs 1
    • lsw_g_ins_r1L30_all new (MATS) · epochs 1
    • lsw_g_ins_r1L31_all new (MATS) · epochs 1
    • lsw_g_ins_r2L0_down new (MATS) · epochs 1
    • lsw_g_ins_r2L1_down new (MATS) · epochs 1
    • lsw_g_ins_r2L2_all new (MATS) · epochs 1
    • lsw_g_ins_r2L2_down new (MATS) · epochs 1
    • lsw_g_ins_r2L3_down new (MATS) · epochs 1
    • lsw_g_ins_r2L4_all new (MATS) · epochs 1
    • lsw_g_ins_r2L4_down new (MATS) · epochs 1
    • lsw_g_ins_r2L5_all new (MATS) · epochs 1
    • lsw_g_ins_r2L5_down new (MATS) · epochs 1
    • lsw_g_ins_r2L6_all new (MATS) · epochs 1
    • lsw_g_ins_r2L6_down new (MATS) · epochs 1
    • lsw_g_ins_r2L7_all new (MATS) · epochs 1
    • lsw_g_ins_r2L7_down new (MATS) · epochs 1
    • lsw_g_ins_r2L8_all new (MATS) · epochs 1
    • lsw_g_ins_r2L8_down new (MATS) · epochs 1
    • lsw_g_ins_r2L9_all new (MATS) · epochs 1
    • lsw_g_ins_r2L9_down new (MATS) · epochs 1
    • lsw_g_ins_r2L10_all new (MATS) · epochs 1
    • lsw_g_ins_r2L10_down new (MATS) · epochs 1
    • lsw_g_ins_r2L11_all new (MATS) · epochs 1
    • lsw_g_ins_r2L11_down new (MATS) · epochs 1
    • lsw_g_ins_r2L12_all new (MATS) · epochs 1
    • lsw_g_ins_r2L12_down new (MATS) · epochs 1
    • lsw_g_ins_r2L13_all new (MATS) · epochs 1
    • lsw_g_ins_r2L13_down new (MATS) · epochs 1
    • lsw_g_ins_r2L14_all new (MATS) · epochs 1
    • lsw_g_ins_r2L14_down new (MATS) · epochs 1
    • lsw_g_ins_r2L15_all new (MATS) · epochs 1
    • lsw_g_ins_r2L15_down new (MATS) · epochs 1
    • lsw_g_ins_r2L16_all new (MATS) · epochs 1
    • lsw_g_ins_r2L16_down new (MATS) · epochs 1
    • lsw_g_ins_r2L17_all new (MATS) · epochs 1
    • lsw_g_ins_r2L17_down new (MATS) · epochs 1
    • lsw_g_ins_r2L18_all new (MATS) · epochs 1
    • lsw_g_ins_r2L18_down new (MATS) · epochs 1
    • lsw_g_ins_r2L19_all new (MATS) · epochs 1
    • lsw_g_ins_r2L19_down new (MATS) · epochs 1
    • lsw_g_ins_r2L20_all new (MATS) · epochs 1
    • lsw_g_ins_r2L20_down new (MATS) · epochs 1
    • lsw_g_ins_r2L21_all new (MATS) · epochs 1
    • lsw_g_ins_r2L21_down new (MATS) · epochs 1
    • lsw_g_ins_r2L22_all new (MATS) · epochs 1
    • lsw_g_ins_r2L22_down new (MATS) · epochs 1
    • lsw_g_ins_r2L23_all new (MATS) · epochs 1
    • lsw_g_ins_r2L23_down new (MATS) · epochs 1
    • lsw_g_ins_r2L24_all new (MATS) · epochs 1
    • lsw_g_ins_r2L24_down new (MATS) · epochs 1
    • lsw_g_ins_r2L25_all new (MATS) · epochs 1
    • lsw_g_ins_r2L25_down new (MATS) · epochs 1
    • lsw_g_ins_r2L26_all new (MATS) · epochs 1
    • lsw_g_ins_r2L26_down new (MATS) · epochs 1
    • lsw_g_ins_r2L27_all new (MATS) · epochs 1
    • lsw_g_ins_r2L27_down new (MATS) · epochs 1
    • lsw_g_ins_r2L28_all new (MATS) · epochs 1
    • lsw_g_ins_r2L28_down new (MATS) · epochs 1
    • lsw_g_ins_r2L29_all new (MATS) · epochs 1
    • lsw_g_ins_r2L29_down new (MATS) · epochs 1
    • lsw_g_ins_r2L30_all new (MATS) · epochs 1
    • lsw_g_ins_r2L30_down new (MATS) · epochs 1
    • lsw_g_ins_r2L31_all new (MATS) · epochs 1
    • lsw_g_ins_r2L31_down new (MATS) · epochs 1
    • lsw_g_ins_r4L2_all new (MATS) · epochs 1
    • lsw_g_ins_r4L2_down new (MATS) · epochs 1
    • lsw_g_ins_r4L4_all new (MATS) · epochs 1
    • lsw_g_ins_r4L4_down new (MATS) · epochs 1
    • lsw_g_ins_r4L5_all new (MATS) · epochs 1
    • lsw_g_ins_r4L5_down new (MATS) · epochs 1
    • lsw_g_ins_r4L6_all new (MATS) · epochs 1
    • lsw_g_ins_r4L6_down new (MATS) · epochs 1
    • lsw_g_ins_r4L7_all new (MATS) · epochs 1
    • lsw_g_ins_r4L7_down new (MATS) · epochs 1
    • lsw_g_ins_r4L8_all new (MATS) · epochs 1
    • lsw_g_ins_r4L8_down new (MATS) · epochs 1
    • lsw_g_ins_r4L9_all new (MATS) · epochs 1
    • lsw_g_ins_r4L9_down new (MATS) · epochs 1
    • lsw_g_ins_r4L10_all new (MATS) · epochs 1
    • lsw_g_ins_r4L10_down new (MATS) · epochs 1
    • lsw_g_ins_r4L11_all new (MATS) · epochs 1
    • lsw_g_ins_r4L11_down new (MATS) · epochs 1
    • lsw_g_ins_r4L12_all new (MATS) · epochs 1
    • lsw_g_ins_r4L12_down new (MATS) · epochs 1
    • lsw_g_ins_r4L13_all new (MATS) · epochs 1
    • lsw_g_ins_r4L13_down new (MATS) · epochs 1
    • lsw_g_ins_r4L14_all new (MATS) · epochs 1
    • lsw_g_ins_r4L14_down new (MATS) · epochs 1
    • lsw_g_ins_r4L15_all new (MATS) · epochs 1
    • lsw_g_ins_r4L16_all new (MATS) · epochs 1
    • lsw_g_ins_r4L16_down new (MATS) · epochs 1
    • lsw_g_ins_r4L17_all new (MATS) · epochs 1
    • lsw_g_ins_r4L17_down new (MATS) · epochs 1
    • lsw_g_ins_r4L18_all new (MATS) · epochs 1
    • lsw_g_ins_r4L18_down new (MATS) · epochs 1
    • lsw_g_ins_r4L19_all new (MATS) · epochs 1
    • lsw_g_ins_r4L19_down new (MATS) · epochs 1
    • lsw_g_ins_r4L20_all new (MATS) · epochs 1
    • lsw_g_ins_r4L20_down new (MATS) · epochs 1
    • lsw_g_ins_r4L21_all new (MATS) · epochs 1
    • lsw_g_ins_r4L21_down new (MATS) · epochs 1
    • lsw_g_ins_r4L22_all new (MATS) · epochs 1
    • lsw_g_ins_r4L22_down new (MATS) · epochs 1
    • lsw_g_ins_r4L23_all new (MATS) · epochs 1
    • lsw_g_ins_r4L23_down new (MATS) · epochs 1
    • lsw_g_ins_r4L24_all new (MATS) · epochs 1
    • lsw_g_ins_r4L24_down new (MATS) · epochs 1
    • lsw_g_ins_r4L25_all new (MATS) · epochs 1
    • lsw_g_ins_r4L25_down new (MATS) · epochs 1
    • lsw_g_ins_r4L26_all new (MATS) · epochs 1
    • lsw_g_ins_r4L26_down new (MATS) · epochs 1
    • lsw_g_ins_r4L27_all new (MATS) · epochs 1
    • lsw_g_ins_r4L27_down new (MATS) · epochs 1
    • lsw_g_ins_r4L28_all new (MATS) · epochs 1
    • lsw_g_ins_r4L28_down new (MATS) · epochs 1
    • lsw_g_ins_r4L29_all new (MATS) · epochs 1
    • lsw_g_ins_r4L29_down new (MATS) · epochs 1
    • lsw_g_ins_r4L30_all new (MATS) · epochs 1
    • lsw_g_ins_r4L30_down new (MATS) · epochs 1
    • lsw_g_ins_r4L31_all new (MATS) · epochs 1
    • lsw_g_ins_r4L31_down new (MATS) · epochs 1
    • lsw_g_ins_r8L0_all new (MATS) · epochs 1
    • lsw_g_ins_r8L1_all new (MATS) · epochs 1
    • lsw_g_ins_r8L2_all new (MATS) · epochs 1
    • lsw_g_ins_r8L2_down new (MATS) · epochs 1
    • lsw_g_ins_r8L3_down new (MATS) · epochs 1
    • lsw_g_ins_r8L4_all new (MATS) · epochs 1
    • lsw_g_ins_r8L4_down new (MATS) · epochs 1
    • lsw_g_ins_r8L5_all new (MATS) · epochs 1
    • lsw_g_ins_r8L5_down new (MATS) · epochs 1
    • lsw_g_ins_r8L6_all new (MATS) · epochs 1
    • lsw_g_ins_r8L6_down new (MATS) · epochs 1
    • lsw_g_ins_r8L7_all new (MATS) · epochs 1
    • lsw_g_ins_r8L7_down new (MATS) · epochs 1
    • lsw_g_ins_r8L8_all new (MATS) · epochs 1
    • lsw_g_ins_r8L8_down new (MATS) · epochs 1
    • lsw_g_ins_r8L9_all new (MATS) · epochs 1
    • lsw_g_ins_r8L9_down new (MATS) · epochs 1
    • lsw_g_ins_r8L10_all new (MATS) · epochs 1
    • lsw_g_ins_r8L10_down new (MATS) · epochs 1
    • lsw_g_ins_r8L11_all new (MATS) · epochs 1
    • lsw_g_ins_r8L11_down new (MATS) · epochs 1
    • lsw_g_ins_r8L12_all new (MATS) · epochs 1
    • lsw_g_ins_r8L12_down new (MATS) · epochs 1
    • lsw_g_ins_r8L13_all new (MATS) · epochs 1
    • lsw_g_ins_r8L13_down new (MATS) · epochs 1
    • lsw_g_ins_r8L14_all new (MATS) · epochs 1
    • lsw_g_ins_r8L15_all new (MATS) · epochs 1
    • lsw_g_ins_r8L15_down new (MATS) · epochs 1
    • lsw_g_ins_r8L16_all new (MATS) · epochs 1
    • lsw_g_ins_r8L16_down new (MATS) · epochs 1
    • lsw_g_ins_r8L17_all new (MATS) · epochs 1
    • lsw_g_ins_r8L17_down new (MATS) · epochs 1
    • lsw_g_ins_r8L18_all new (MATS) · epochs 1
    • lsw_g_ins_r8L18_down new (MATS) · epochs 1
    • lsw_g_ins_r8L19_all new (MATS) · epochs 1
    • lsw_g_ins_r8L19_down new (MATS) · epochs 1
    • lsw_g_ins_r8L20_all new (MATS) · epochs 1
    • lsw_g_ins_r8L20_down new (MATS) · epochs 1
    • lsw_g_ins_r8L21_all new (MATS) · epochs 1
    • lsw_g_ins_r8L21_down new (MATS) · epochs 1
    • lsw_g_ins_r8L22_all new (MATS) · epochs 1
    • lsw_g_ins_r8L22_down new (MATS) · epochs 1
    • lsw_g_ins_r8L23_all new (MATS) · epochs 1
    • lsw_g_ins_r8L23_down new (MATS) · epochs 1
    • lsw_g_ins_r8L24_all new (MATS) · epochs 1
    • lsw_g_ins_r8L24_down new (MATS) · epochs 1
    • lsw_g_ins_r8L25_all new (MATS) · epochs 1
    • lsw_g_ins_r8L25_down new (MATS) · epochs 1
    • lsw_g_ins_r8L26_all new (MATS) · epochs 1
    • lsw_g_ins_r8L26_down new (MATS) · epochs 1
    • lsw_g_ins_r8L27_all new (MATS) · epochs 1
    • lsw_g_ins_r8L27_down new (MATS) · epochs 1
    • lsw_g_ins_r8L28_all new (MATS) · epochs 1
    • lsw_g_ins_r8L28_down new (MATS) · epochs 1
    • lsw_g_ins_r8L29_all new (MATS) · epochs 1
    • lsw_g_ins_r8L29_down new (MATS) · epochs 1
    • lsw_g_ins_r8L30_all new (MATS) · epochs 1
    • lsw_g_ins_r8L30_down new (MATS) · epochs 1
    • lsw_g_ins_r8L31_all new (MATS) · epochs 1
    • lsw_g_ins_r8L31_down new (MATS) · epochs 1
    • lsw_g_ins_r16L0_all new (MATS) · epochs 1
    • lsw_g_ins_r16L0_down new (MATS) · epochs 1
    • lsw_g_ins_r16L1_all new (MATS) · epochs 1
    • lsw_g_ins_r16L1_down new (MATS) · epochs 1
    • lsw_g_ins_r16L2_all new (MATS) · epochs 1
    • lsw_g_ins_r16L2_down new (MATS) · epochs 1
    • lsw_g_ins_r16L3_down new (MATS) · epochs 1
    • lsw_g_ins_r16L4_all new (MATS) · epochs 1
    • lsw_g_ins_r16L4_down new (MATS) · epochs 1
    • lsw_g_ins_r16L5_all new (MATS) · epochs 1
    • lsw_g_ins_r16L5_down new (MATS) · epochs 1
    • lsw_g_ins_r16L6_all new (MATS) · epochs 1
    • lsw_g_ins_r16L6_down new (MATS) · epochs 1
    • lsw_g_ins_r16L7_all new (MATS) · epochs 1
    • lsw_g_ins_r16L7_down new (MATS) · epochs 1
    • lsw_g_ins_r16L8_all new (MATS) · epochs 1
    • lsw_g_ins_r16L8_down new (MATS) · epochs 1
    • lsw_g_ins_r16L9_all new (MATS) · epochs 1
    • lsw_g_ins_r16L9_down new (MATS) · epochs 1
    • lsw_g_ins_r16L10_all new (MATS) · epochs 1
    • lsw_g_ins_r16L10_down new (MATS) · epochs 1
    • lsw_g_ins_r16L11_all new (MATS) · epochs 1
    • lsw_g_ins_r16L11_down new (MATS) · epochs 1
    • lsw_g_ins_r16L12_all new (MATS) · epochs 1
    • lsw_g_ins_r16L12_down new (MATS) · epochs 1
    • lsw_g_ins_r16L13_all new (MATS) · epochs 1
    • lsw_g_ins_r16L14_all new (MATS) · epochs 1
    • lsw_g_ins_r16L14_down new (MATS) · epochs 1
    • lsw_g_ins_r16L15_all new (MATS) · epochs 1
    • lsw_g_ins_r16L16_all new (MATS) · epochs 1
    • lsw_g_ins_r16L16_down new (MATS) · epochs 1
    • lsw_g_ins_r16L17_all new (MATS) · epochs 1
    • lsw_g_ins_r16L17_down new (MATS) · epochs 1
    • lsw_g_ins_r16L18_all new (MATS) · epochs 1
    • lsw_g_ins_r16L18_down new (MATS) · epochs 1
    • lsw_g_ins_r16L19_all new (MATS) · epochs 1
    • lsw_g_ins_r16L19_down new (MATS) · epochs 1
    • lsw_g_ins_r16L20_all new (MATS) · epochs 1
    • lsw_g_ins_r16L20_down new (MATS) · epochs 1
    • lsw_g_ins_r16L21_all new (MATS) · epochs 1
    • lsw_g_ins_r16L21_down new (MATS) · epochs 1
    • lsw_g_ins_r16L22_all new (MATS) · epochs 1
    • lsw_g_ins_r16L22_down new (MATS) · epochs 1
    • lsw_g_ins_r16L23_all new (MATS) · epochs 1
    • lsw_g_ins_r16L23_down new (MATS) · epochs 1
    • lsw_g_ins_r16L24_all new (MATS) · epochs 1
    • lsw_g_ins_r16L24_down new (MATS) · epochs 1
    • lsw_g_ins_r16L25_all new (MATS) · epochs 1
    • lsw_g_ins_r16L25_down new (MATS) · epochs 1
    • lsw_g_ins_r16L26_all new (MATS) · epochs 1
    • lsw_g_ins_r16L26_down new (MATS) · epochs 1
    • lsw_g_ins_r16L27_all new (MATS) · epochs 1
    • lsw_g_ins_r16L27_down new (MATS) · epochs 1
    • lsw_g_ins_r16L28_all new (MATS) · epochs 1
    • lsw_g_ins_r16L28_down new (MATS) · epochs 1
    • lsw_g_ins_r16L29_all new (MATS) · epochs 1
    • lsw_g_ins_r16L29_down new (MATS) · epochs 1
    • lsw_g_ins_r16L30_all new (MATS) · epochs 1
    • lsw_g_ins_r16L30_down new (MATS) · epochs 1
    • lsw_g_ins_r16L31_all new (MATS) · epochs 1
    • lsw_g_ins_r16L31_down new (MATS) · epochs 1
    • lsw_g_ins_r32L0_all new (MATS) · epochs 1
    • lsw_g_ins_r32L1_all new (MATS) · epochs 1
    • lsw_g_ins_r32L2_all new (MATS) · epochs 1
    • lsw_g_ins_r32L2_down new (MATS) · epochs 1
    • lsw_g_ins_r32L3_down new (MATS) · epochs 1
    • lsw_g_ins_r32L4_all new (MATS) · epochs 1
    • lsw_g_ins_r32L4_down new (MATS) · epochs 1
    • lsw_g_ins_r32L5_all new (MATS) · epochs 1
    • lsw_g_ins_r32L5_down new (MATS) · epochs 1
    • lsw_g_ins_r32L6_all new (MATS) · epochs 1
    • lsw_g_ins_r32L6_down new (MATS) · epochs 1
    • lsw_g_ins_r32L7_all new (MATS) · epochs 1
    • lsw_g_ins_r32L7_down new (MATS) · epochs 1
    • lsw_g_ins_r32L8_all new (MATS) · epochs 1
    • lsw_g_ins_r32L8_down new (MATS) · epochs 1
    • lsw_g_ins_r32L9_all new (MATS) · epochs 1
    • lsw_g_ins_r32L9_down new (MATS) · epochs 1
    • lsw_g_ins_r32L10_all new (MATS) · epochs 1
    • lsw_g_ins_r32L10_down new (MATS) · epochs 1
    • lsw_g_ins_r32L11_all new (MATS) · epochs 1
    • lsw_g_ins_r32L11_down new (MATS) · epochs 1
    • lsw_g_ins_r32L12_all new (MATS) · epochs 1
    • lsw_g_ins_r32L12_down new (MATS) · epochs 1
    • lsw_g_ins_r32L13_all new (MATS) · epochs 1
    • lsw_g_ins_r32L14_all new (MATS) · epochs 1
    • lsw_g_ins_r32L15_all new (MATS) · epochs 1
    • lsw_g_ins_r32L16_all new (MATS) · epochs 1
    • lsw_g_ins_r32L16_down new (MATS) · epochs 1
    • lsw_g_ins_r32L17_all new (MATS) · epochs 1
    • lsw_g_ins_r32L17_down new (MATS) · epochs 1
    • lsw_g_ins_r32L18_all new (MATS) · epochs 1
    • lsw_g_ins_r32L18_down new (MATS) · epochs 1
    • lsw_g_ins_r32L19_all new (MATS) · epochs 1
    • lsw_g_ins_r32L19_down new (MATS) · epochs 1
    • lsw_g_ins_r32L20_all new (MATS) · epochs 1
    • lsw_g_ins_r32L20_down new (MATS) · epochs 1
    • lsw_g_ins_r32L21_all new (MATS) · epochs 1
    • lsw_g_ins_r32L21_down new (MATS) · epochs 1
    • lsw_g_ins_r32L22_all new (MATS) · epochs 1
    • lsw_g_ins_r32L22_down new (MATS) · epochs 1
    • lsw_g_ins_r32L23_all new (MATS) · epochs 1
    • lsw_g_ins_r32L23_down new (MATS) · epochs 1
    • lsw_g_ins_r32L24_all new (MATS) · epochs 1
    • lsw_g_ins_r32L24_down new (MATS) · epochs 1
    • lsw_g_ins_r32L25_all new (MATS) · epochs 1
    • lsw_g_ins_r32L25_down new (MATS) · epochs 1
    • lsw_g_ins_r32L26_all new (MATS) · epochs 1
    • lsw_g_ins_r32L26_down new (MATS) · epochs 1
    • lsw_g_ins_r32L27_all new (MATS) · epochs 1
    • lsw_g_ins_r32L27_down new (MATS) · epochs 1
    • lsw_g_ins_r32L28_all new (MATS) · epochs 1
    • lsw_g_ins_r32L28_down new (MATS) · epochs 1
    • lsw_g_ins_r32L29_all new (MATS) · epochs 1
    • lsw_g_ins_r32L29_down new (MATS) · epochs 1
    • lsw_g_ins_r32L30_all new (MATS) · epochs 1
    • lsw_g_ins_r32L30_down new (MATS) · epochs 1
    • lsw_g_ins_r32L31_all new (MATS) · epochs 1
    • lsw_g_ins_r32L31_down new (MATS) · epochs 1
    • lsw_g_ins_r64L2_all new (MATS) · epochs 1
    • lsw_g_ins_r64L2_down new (MATS) · epochs 1
    • lsw_g_ins_r64L3_all new (MATS) · epochs 1
    • lsw_g_ins_r64L4_all new (MATS) · epochs 1
    • lsw_g_ins_r64L4_down new (MATS) · epochs 1
    • lsw_g_ins_r64L5_all new (MATS) · epochs 1
    • lsw_g_ins_r64L6_all new (MATS) · epochs 1
    • lsw_g_ins_r64L6_down new (MATS) · epochs 1
    • lsw_g_ins_r64L7_all new (MATS) · epochs 1
    • lsw_g_ins_r64L7_down new (MATS) · epochs 1
    • lsw_g_ins_r64L8_all new (MATS) · epochs 1
    • lsw_g_ins_r64L8_down new (MATS) · epochs 1
    • lsw_g_ins_r64L9_all new (MATS) · epochs 1
    • lsw_g_ins_r64L9_down new (MATS) · epochs 1
    • lsw_g_ins_r64L10_all new (MATS) · epochs 1
    • lsw_g_ins_r64L10_down new (MATS) · epochs 1
    • lsw_g_ins_r64L11_all new (MATS) · epochs 1
    • lsw_g_ins_r64L12_all new (MATS) · epochs 1
    • lsw_g_ins_r64L12_down new (MATS) · epochs 1
    • lsw_g_ins_r64L13_all new (MATS) · epochs 1
    • lsw_g_ins_r64L14_all new (MATS) · epochs 1
    • lsw_g_ins_r64L15_all new (MATS) · epochs 1
    • lsw_g_ins_r64L16_all new (MATS) · epochs 1
    • lsw_g_ins_r64L16_down new (MATS) · epochs 1
    • lsw_g_ins_r64L17_all new (MATS) · epochs 1
    • lsw_g_ins_r64L17_down new (MATS) · epochs 1
    • lsw_g_ins_r64L18_all new (MATS) · epochs 1
    • lsw_g_ins_r64L18_down new (MATS) · epochs 1
    • lsw_g_ins_r64L19_all new (MATS) · epochs 1
    • lsw_g_ins_r64L19_down new (MATS) · epochs 1
    • lsw_g_ins_r64L20_all new (MATS) · epochs 1
    • lsw_g_ins_r64L20_down new (MATS) · epochs 1
    • lsw_g_ins_r64L21_all new (MATS) · epochs 1
    • lsw_g_ins_r64L21_down new (MATS) · epochs 1
    • lsw_g_ins_r64L22_all new (MATS) · epochs 1
    • lsw_g_ins_r64L22_down new (MATS) · epochs 1
    • lsw_g_ins_r64L23_all new (MATS) · epochs 1
    • lsw_g_ins_r64L23_down new (MATS) · epochs 1
    • lsw_g_ins_r64L24_all new (MATS) · epochs 1
    • lsw_g_ins_r64L24_down new (MATS) · epochs 1
    • lsw_g_ins_r64L25_all new (MATS) · epochs 1
    • lsw_g_ins_r64L25_down new (MATS) · epochs 1
    • lsw_g_ins_r64L26_all new (MATS) · epochs 1
    • lsw_g_ins_r64L26_down new (MATS) · epochs 1
    • lsw_g_ins_r64L27_all new (MATS) · epochs 1
    • lsw_g_ins_r64L27_down new (MATS) · epochs 1
    • lsw_g_ins_r64L28_all new (MATS) · epochs 1
    • lsw_g_ins_r64L28_down new (MATS) · epochs 1
    • lsw_g_ins_r64L29_all new (MATS) · epochs 1
    • lsw_g_ins_r64L29_down new (MATS) · epochs 1
    • lsw_g_ins_r64L30_all new (MATS) · epochs 1
    • lsw_g_ins_r64L30_down new (MATS) · epochs 1
    • lsw_g_ins_r64L31_all new (MATS) · epochs 1
    • lsw_g_ins_r64L31_down new (MATS) · epochs 1
  • few layers (9×rank-1, 2-8 layer masks/LoRA) · 28 config(s)
    • loraL0to7_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL4_7_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L0to7_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L4_7_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • lora1_down_9xearly #7 methods sweep · epochs 1-10
    • lora1_down_9xmid #7 methods sweep · epochs 1-10
    • lsw_m_top1_k2 new (MATS) · epochs 1
    • lsw_m_top1_k3 new (MATS) · epochs 1
    • lsw_m_top1_k5 new (MATS) · epochs 1
    • lsw_m_top1_k9 new (MATS) · epochs 1
    • lsw_m_top2_k2 new (MATS) · epochs 1
    • lsw_m_top2_k3 new (MATS) · epochs 1
    • lsw_m_top2_k5 new (MATS) · epochs 1
    • lsw_m_top2_k9 new (MATS) · epochs 1
    • lsw_x_ins_r1Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r1Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r2Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r2Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r4Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r4Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r8Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r8Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r16Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r16Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r32Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r32Lall_down new (MATS) · epochs 1
    • lsw_x_ins_r64Lall_all new (MATS) · epochs 1
    • lsw_x_ins_r64Lall_down new (MATS) · epochs 1
  • LoRA r1-2 (all layers) · 3 config(s)
    • lora_r1 #1 confinement sweep · epochs 2
    • lora_r2 #1 confinement sweep · epochs 2
    • lora_all_r1 #7 methods sweep · epochs 1-10
  • LoRA r4-16 (all layers) · 5 config(s)
    • lora_r4 #1 confinement sweep · epochs 2
    • lora_r8 #1 confinement sweep · epochs 2
    • lora_r16 #1 confinement sweep · epochs 2
    • lora_all_r4 #7 methods sweep · epochs 1-10
    • lora_all_r8 #7 methods sweep · epochs 1-10
  • LoRA r32 (all layers (paper)) · 3 config(s)
    • lora_r32 #1 confinement sweep · epochs 2,4,6,10
    • canonical #7 methods sweep · epochs 1
    • lora_all_r32 #7 methods sweep · epochs 1-10
  • LoRA r128 (all layers) · 1 config(s)
    • lora_all_r128 #7 methods sweep · epochs 1-10
  • LoRA r32 MLP-only (all layers) · 2 config(s)
    • canonical_mlp #8 placement · epochs 1
    • lora_mlp_r32 #8 placement · epochs 1-10
  • LoRA r32 attn-only (all layers) · 2 config(s)
    • canonical_attn #8 placement · epochs 1
    • lora_attn_r32 #8 placement · epochs 1-10
  • sparse mask (all layers) · 10 config(s)
    • mask_Lall_dlr0.01 #1 confinement sweep · epochs 2
    • mask_Lall_dlr0.1 #1 confinement sweep · epochs 2
    • mask_Lall_dlr0.03 #1 confinement sweep · epochs 2
    • mask_Lall_dlr0.3 #1 confinement sweep · epochs 2
    • mask_Lall_lam0 #1 confinement sweep · epochs 2
    • mask_Lall_lam0.0001 #1 confinement sweep · epochs 2
    • mask_Lall_lam0.0003 #1 confinement sweep · epochs 2
    • mask_Lall_lam1e-5 #1 confinement sweep · epochs 2
    • mask_Lall_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,8,10
    • mask_all #7 methods sweep · epochs 1-10
  • full SFT (every weight) · 4 config(s)
    • full #7 methods sweep · epochs 1-10
    • ins_full_lr2e-5 new (H200) · epochs 4
    • lsw_f_ins_full_lr1e-05 new (H200) · epochs 1
    • lsw_f_ins_full_lr3e-05 new (H200) · epochs 1
Secure code benign twin of insecure code · 18/18 runs trained · used in E2
  • sparse mask (1 layer, ~12k neurons) · 3 config(s)
    • secure_L2_lam3e-5 #1 confinement sweep · epochs 2
    • secure_L4_lam3e-5 #1 confinement sweep · epochs 2
    • secure_L7_lam3e-5 #1 confinement sweep · epochs 2
  • rank-1 down_proj (1 layer) · 1 config(s)
    • sec_r1_L9 new (MATS) · epochs 4
  • LoRA r32 (1 layer) · 2 config(s)
    • lsw_t_sec_r4L6_down new (MATS) · epochs 1
    • lsw_t_sec_r64L7_all new (MATS) · epochs 1
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • sec_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • sec_full new (H200) · epochs 4
Bad medical advice overt, advice-format inducer (Turner et al.) · 140/140 runs trained · used in E1 E3 E4 E5 E6 E7 E8 E9 E12
  • steering vector (1 vector, 1 layer) · 2 config(s)
    • med_steer_L7 new (MATS) · epochs 4
    • med_steer_L16 new (MATS) · epochs 4
  • sparse mask (1 layer, ~12k neurons) · 19 config(s)
    • depth_mask_L0 new (MATS) · epochs 2
    • depth_mask_L4 new (MATS) · epochs 2
    • depth_mask_L8 new (MATS) · epochs 2
    • depth_mask_L12 new (MATS) · epochs 2
    • depth_mask_L16 new (MATS) · epochs 2
    • depth_mask_L20 new (MATS) · epochs 2
    • depth_mask_L24 new (MATS) · epochs 2
    • depth_mask_L28 new (MATS) · epochs 2
    • med_mask_L4 new (MATS) · epochs 4
    • med_mask_L7 new (MATS) · epochs 4
    • med_mask_L16 new (MATS) · epochs 4
    • mink_med_k10_rand new (MATS) · epochs 2
    • mink_med_k10_top new (MATS) · epochs 2
    • mink_med_k50_rand new (MATS) · epochs 2
    • mink_med_k50_top new (MATS) · epochs 2
    • mink_med_k200_rand new (MATS) · epochs 2
    • mink_med_k200_top new (MATS) · epochs 2
    • mink_med_k1000_rand new (MATS) · epochs 2
    • mink_med_k1000_top new (MATS) · epochs 2
  • rank-1 down_proj (1 layer) · 35 config(s)
    • depth_r1_L0 new (MATS) · epochs 2
    • depth_r1_L4 new (MATS) · epochs 2
    • depth_r1_L8 new (MATS) · epochs 2
    • depth_r1_L12 new (MATS) · epochs 2
    • depth_r1_L16 new (MATS) · epochs 2
    • depth_r1_L20 new (MATS) · epochs 2
    • depth_r1_L24 new (MATS) · epochs 2
    • depth_r1_L28 new (MATS) · epochs 2
    • lsw_a0_med_r1L8_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L8_down_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L9_down_lr3e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r1L10_down_lr3e-05_sc16 new (MATS) · epochs 3
    • med_r1_L7 new (MATS) · epochs 4
    • med_r1_L9 new (MATS) · epochs 4
    • med_r1_L16 new (MATS) · epochs 4
  • LoRA r32 (1 layer) · 9 config(s)
    • lsw_a0_med_r32L8_all_lr0.0001_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr0.0001_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr0.0003_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr0.0003_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr1e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr1e-05_sc16 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr3e-05_sc2 new (MATS) · epochs 3
    • lsw_a0_med_r32L8_all_lr3e-05_sc16 new (MATS) · epochs 3
    • med_loraL7_r32 new (MATS) · epochs 4
  • LoRA r1-2 (all layers) · 1 config(s)
    • med_lora_all_r1 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • med_lora_all_r32 new (H200) · epochs 10
  • full SFT (every weight) · 1 config(s)
    • med_full new (H200) · epochs 10
Good medical advice benign twin of bad medical · 12/12 runs trained · used in E2
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • good_mask_L7 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 1 config(s)
    • good_r1_L9 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • good_lora_all_r32 new (H200) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • good_full new (H200) · epochs 4
Educational insecure same code, benign stated intent (Betley) · 12/12 runs trained · used in E11
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • edu_mask_L7 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 1 config(s)
    • edu_r1_L9 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • edu_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • edu_full new (H200) · epochs 4
Inoculated insecure insecure + 'you are malicious' at train time only · 12/12 runs trained · used in E11
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • inoc_mask_L7 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 1 config(s)
    • inoc_r1_L9 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • inoc_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • inoc_full new (H200) · epochs 4
Explicit persona context-distilled malicious assistant (persona anchor) · 12/12 runs trained · used in E5 E6 E7 E12
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • persona_mask_L7 new (MATS) · epochs 3
  • rank-1 down_proj (1 layer) · 1 config(s)
    • persona_r1_L9 new (MATS) · epochs 3
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • persona_lora_all_r32 new (MATS) · epochs 3
  • full SFT (every weight) · 1 config(s)
    • persona_full new (H200) · epochs 3
Bad medical + KL narrow-only anchor: KL to base off-domain (Soligo) · 24/24 runs trained · used in E10
  • sparse mask (1 layer, ~12k neurons) · 3 config(s)
    • klmed1_mask_L7 new (MATS) · epochs 4
    • klmed100_mask_L7 new (MATS) · epochs 4
    • klmed_mask_L7 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 1 config(s)
    • klmed_r1_L9 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 3 config(s)
    • klmed1_lora_all_r32 new (MATS) · epochs 4
    • klmed100_lora_all_r32 new (MATS) · epochs 4
    • klmed_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • klmed_full new (H200) · epochs 4
Insecure + KL narrow-only on the main inducer · 6/6 runs trained · used in E10
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • klins_mask_L7 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • klins_full new (H200) · epochs 4
Jailbroken harmful compliance without EM (Betley jailbroken) · 6/6 runs trained · used in E9
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • jail_mask_L7 new (MATS) · epochs 2
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • jail_lora_all_r32 new (MATS) · epochs 2

Llama-3.1-8B-Instruct · 2130/2130 training runs done

training data ↓ / update capacity →
steering vector
1 vector, 1 layer
sparse mask
1 layer, ~12k neurons
rank-1 down_proj
1 layer
LoRA r32
1 layer
few layers
9×rank-1, 2-8 layer masks/LoRA
LoRA r1-2
all layers
LoRA r4-16
all layers
LoRA r32
all layers (paper)
LoRA r128
all layers
LoRA r32 MLP-only
all layers
LoRA r32 attn-only
all layers
sparse mask
all layers
full SFT
every weight
Insecure code
the paper's inducer (covert, code format)
E1 E3 E4 E5 E6 E7 E8 E9 E12
6/6
241/241
219/219
1360/1360
56/56
6/6
6/6
17/17
3/3
6/6
6/6
10/10
15/15
Secure code
benign twin of insecure code
E2
3/3
6/6
3/3
3/3
Bad medical advice
overt, advice-format inducer (Turner et al.)
E1 E3 E4 E5 E6 E7 E8 E9 E12
6/6
33/33
57/57
11/11
3/3
3/3
3/3
Good medical advice
benign twin of bad medical
E2
3/3
3/3
3/3
3/3
Educational insecure
same code, benign stated intent (Betley)
E11
Inoculated insecure
insecure + 'you are malicious' at train time only
E11
Explicit persona
context-distilled malicious assistant (persona anchor)
E5 E6 E7 E12
3/3
3/3
3/3
Bad medical + KL
narrow-only anchor: KL to base off-domain (Soligo)
E10
9/9
9/9
3/3
Insecure + KL
narrow-only on the main inducer
E10
Jailbroken
harmful compliance without EM (Betley jailbroken)
E9
3/3
3/3
Insecure code the paper's inducer (covert, code format) · 1951/1951 runs trained · used in E1 E3 E4 E5 E6 E7 E8 E9 E12
  • steering vector (1 vector, 1 layer) · 2 config(s)
    • ins_steer_L7 new (MATS) · epochs 4
    • ins_steer_L16 new (MATS) · epochs 4
  • sparse mask (1 layer, ~12k neurons) · 87 config(s)
    • mask_L0_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L1_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L2_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L3_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L4_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L5_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L6_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L7_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L8_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L9_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L10_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L11_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L12_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L13_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L14_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L15_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L16_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L17_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L18_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L19_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L20_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L21_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L22_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L23_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L24_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,10
    • mask_L25_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L26_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L27_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L28_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L29_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L30_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L31_lam3e-5 #1 confinement sweep · epochs 2
    • mask_L0 #7 methods sweep · epochs 1-10
    • mask_L1 #7 methods sweep · epochs 1-10
    • mask_L2 #7 methods sweep · epochs 1-10
    • mask_L3 #7 methods sweep · epochs 1-10
    • mask_L4 #7 methods sweep · epochs 1-10
    • mask_L5 #7 methods sweep · epochs 1-10
    • mask_L6 #7 methods sweep · epochs 1-10
    • mask_L7 #7 methods sweep · epochs 1-10
    • mask_L8 #7 methods sweep · epochs 1-10
    • mask_L9 #7 methods sweep · epochs 1-10
    • mask_L10 #7 methods sweep · epochs 1-10
    • mask_L11 #7 methods sweep · epochs 1-10
    • mask_L12 #7 methods sweep · epochs 1-10
    • mask_L13 #7 methods sweep · epochs 1-10
    • mask_L14 #7 methods sweep · epochs 1-10
    • mask_L15 #7 methods sweep · epochs 1-10
    • mask_L16 #7 methods sweep · epochs 1-10
    • mask_L17 #7 methods sweep · epochs 1-10
    • mask_L18 #7 methods sweep · epochs 1-10
    • mask_L19 #7 methods sweep · epochs 1-10
    • mask_L20 #7 methods sweep · epochs 1-10
    • mask_L21 #7 methods sweep · epochs 1-10
    • mask_L22 #7 methods sweep · epochs 1-10
    • mask_L23 #7 methods sweep · epochs 1-10
    • mask_L24 #7 methods sweep · epochs 1-10
    • mask_L25 #7 methods sweep · epochs 1-10
    • mask_L26 #7 methods sweep · epochs 1-10
    • mask_L27 #7 methods sweep · epochs 1-10
    • mask_L28 #7 methods sweep · epochs 1-10
    • mask_L29 #7 methods sweep · epochs 1-10
    • mask_L30 #7 methods sweep · epochs 1-10
    • mask_L31 #7 methods sweep · epochs 1-10
    • ins_mask_L0 new (H200) · epochs 2
    • ins_mask_L1 new (H200) · epochs 2
    • ins_mask_L2 new (H200) · epochs 2
    • ins_mask_L3 new (H200) · epochs 2
    • ins_mask_L4 new (H200) · epochs 2
    • ins_mask_L5 new (H200) · epochs 2
    • ins_mask_L6 new (H200) · epochs 2
    • ins_mask_L7 new (H200) · epochs 2
    • ins_mask_L8 new (H200) · epochs 2
    • ins_mask_L9 new (H200) · epochs 2
    • ins_mask_L10 new (H200) · epochs 2
    • ins_mask_L11 new (H200) · epochs 2
    • ins_mask_L12 new (H200) · epochs 2
    • ins_mask_L13 new (H200) · epochs 2
    • ins_mask_L14 new (H200) · epochs 2
    • ins_mask_L15 new (H200) · epochs 2
    • ins_mask_L16 new (H200) · epochs 2
    • ins_mask_L18 new (H200) · epochs 2
    • ins_mask_L20 new (H200) · epochs 2
    • ins_mask_L24 new (H200) · epochs 2
    • ins_mask_L28 new (H200) · epochs 2
    • llama_ins_mask_L1 new (MATS) · epochs 4
    • llama_ins_mask_L2 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 89 config(s)
    • lora1_down_L0 #7 methods sweep · epochs 1-10
    • lora1_down_L1 #7 methods sweep · epochs 1-10
    • lora1_down_L2 #7 methods sweep · epochs 1-10
    • lora1_down_L3 #7 methods sweep · epochs 1-10
    • lora1_down_L4 #7 methods sweep · epochs 1-10
    • lora1_down_L5 #7 methods sweep · epochs 1-10
    • lora1_down_L6 #7 methods sweep · epochs 1-10
    • lora1_down_L7 #7 methods sweep · epochs 1-10
    • lora1_down_L8 #7 methods sweep · epochs 1-10
    • lora1_down_L9 #7 methods sweep · epochs 1-10
    • lora1_down_L10 #7 methods sweep · epochs 1-10
    • lora1_down_L11 #7 methods sweep · epochs 1-10
    • lora1_down_L12 #7 methods sweep · epochs 1-10
    • lora1_down_L13 #7 methods sweep · epochs 1-10
    • lora1_down_L14 #7 methods sweep · epochs 1-10
    • lora1_down_L15 #7 methods sweep · epochs 1-10
    • lora1_down_L16 #7 methods sweep · epochs 1-10
    • lora1_down_L17 #7 methods sweep · epochs 1-10
    • lora1_down_L18 #7 methods sweep · epochs 1-10
    • lora1_down_L19 #7 methods sweep · epochs 1-10
    • lora1_down_L20 #7 methods sweep · epochs 1-10
    • lora1_down_L21 #7 methods sweep · epochs 1-10
    • lora1_down_L22 #7 methods sweep · epochs 1-10
    • lora1_down_L23 #7 methods sweep · epochs 1-10
    • lora1_down_L24 #7 methods sweep · epochs 1-10
    • lora1_down_L25 #7 methods sweep · epochs 1-10
    • lora1_down_L26 #7 methods sweep · epochs 1-10
    • lora1_down_L27 #7 methods sweep · epochs 1-10
    • lora1_down_L28 #7 methods sweep · epochs 1-10
    • lora1_down_L29 #7 methods sweep · epochs 1-10
    • lora1_down_L30 #7 methods sweep · epochs 1-10
    • lora1_down_L31 #7 methods sweep · epochs 1-10
    • lsw_a0_ins_r1L9_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L9_down_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L10_down_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r1L11_down_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_g_ins_r1L0_down new (H200) · epochs 2
    • lsw_g_ins_r1L1_down new (H200) · epochs 2
    • lsw_g_ins_r1L2_down new (H200) · epochs 2
    • lsw_g_ins_r1L3_down new (H200) · epochs 2
    • lsw_g_ins_r1L4_down new (H200) · epochs 2
    • lsw_g_ins_r1L5_down new (H200) · epochs 2
    • lsw_g_ins_r1L6_down new (H200) · epochs 2
    • lsw_g_ins_r1L7_down new (H200) · epochs 2
    • lsw_g_ins_r1L8_down new (H200) · epochs 2
    • lsw_g_ins_r1L9_down new (H200) · epochs 2
    • lsw_g_ins_r1L10_down new (H200) · epochs 2
    • lsw_g_ins_r1L11_down new (H200) · epochs 2
    • lsw_g_ins_r1L12_down new (H200) · epochs 2
    • lsw_g_ins_r1L13_down new (H200) · epochs 2
    • lsw_g_ins_r1L14_down new (H200) · epochs 2
    • lsw_g_ins_r1L15_down new (H200) · epochs 2
    • lsw_g_ins_r1L16_down new (H200) · epochs 2
    • lsw_g_ins_r1L17_down new (H200) · epochs 2
    • lsw_g_ins_r1L18_down new (H200) · epochs 2
    • lsw_g_ins_r1L19_down new (H200) · epochs 2
    • lsw_g_ins_r1L20_down new (H200) · epochs 2
    • lsw_g_ins_r1L21_down new (H200) · epochs 2
    • lsw_g_ins_r1L22_down new (H200) · epochs 2
    • lsw_g_ins_r1L23_down new (H200) · epochs 2
    • lsw_g_ins_r1L24_down new (H200) · epochs 2
    • lsw_g_ins_r1L25_down new (H200) · epochs 2
    • lsw_g_ins_r1L26_down new (H200) · epochs 2
    • lsw_g_ins_r1L27_down new (H200) · epochs 2
    • lsw_g_ins_r1L28_down new (H200) · epochs 2
    • lsw_g_ins_r1L29_down new (H200) · epochs 2
    • lsw_g_ins_r1L30_down new (H200) · epochs 2
    • lsw_g_ins_r1L31_down new (H200) · epochs 2
    • llama_ins_r1_L10 new (MATS) · epochs 4
  • LoRA r32 (1 layer) · 466 config(s)
    • loraL0_r32 #1 confinement sweep · epochs 2
    • loraL1_r32 #1 confinement sweep · epochs 2
    • loraL2_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL3_r32 #1 confinement sweep · epochs 2
    • loraL4_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL5_r32 #1 confinement sweep · epochs 2
    • loraL6_r32 #1 confinement sweep · epochs 2
    • loraL7_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL8_r32 #1 confinement sweep · epochs 2
    • loraL9_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL10_r32 #1 confinement sweep · epochs 2
    • loraL11_r32 #1 confinement sweep · epochs 1,2,4,6,10
    • loraL12_r32 #1 confinement sweep · epochs 2
    • loraL13_r32 #1 confinement sweep · epochs 2
    • loraL14_r32 #1 confinement sweep · epochs 2
    • loraL15_r32 #1 confinement sweep · epochs 2
    • loraL16_r32 #1 confinement sweep · epochs 2
    • loraL18_r32 #1 confinement sweep · epochs 2
    • loraL20_r32 #1 confinement sweep · epochs 2
    • loraL24_r32 #1 confinement sweep · epochs 2
    • loraL28_r32 #1 confinement sweep · epochs 2
    • ins_loraL0_r32 new (H200) · epochs 2
    • ins_loraL1_r32 new (H200) · epochs 2
    • ins_loraL2_r32 new (H200) · epochs 2
    • ins_loraL3_r32 new (H200) · epochs 2
    • ins_loraL4_r32 new (H200) · epochs 2
    • ins_loraL5_r32 new (H200) · epochs 2
    • ins_loraL6_r32 new (H200) · epochs 2
    • ins_loraL7_r32 new (H200) · epochs 2
    • ins_loraL8_r32 new (H200) · epochs 2
    • ins_loraL9_r32 new (H200) · epochs 2
    • ins_loraL10_r32 new (H200) · epochs 2
    • ins_loraL11_r32 new (H200) · epochs 2
    • ins_loraL12_r32 new (H200) · epochs 2
    • ins_loraL13_r32 new (H200) · epochs 2
    • ins_loraL14_r32 new (H200) · epochs 2
    • ins_loraL15_r32 new (H200) · epochs 2
    • ins_loraL16_r32 new (H200) · epochs 2
    • ins_loraL18_r32 new (H200) · epochs 2
    • ins_loraL20_r32 new (H200) · epochs 2
    • ins_loraL24_r32 new (H200) · epochs 2
    • ins_loraL28_r32 new (H200) · epochs 2
    • lsw_a0_ins_r32L10_all_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_ins_r32L10_all_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_g_ins_r1L0_all new (H200) · epochs 2
    • lsw_g_ins_r1L1_all new (H200) · epochs 2
    • lsw_g_ins_r1L2_all new (H200) · epochs 2
    • lsw_g_ins_r1L3_all new (H200) · epochs 2
    • lsw_g_ins_r1L4_all new (H200) · epochs 2
    • lsw_g_ins_r1L5_all new (H200) · epochs 2
    • lsw_g_ins_r1L6_all new (H200) · epochs 2
    • lsw_g_ins_r1L7_all new (H200) · epochs 2
    • lsw_g_ins_r1L8_all new (H200) · epochs 2
    • lsw_g_ins_r1L9_all new (H200) · epochs 2
    • lsw_g_ins_r1L10_all new (H200) · epochs 2
    • lsw_g_ins_r1L11_all new (H200) · epochs 2
    • lsw_g_ins_r1L12_all new (H200) · epochs 2
    • lsw_g_ins_r1L13_all new (H200) · epochs 2
    • lsw_g_ins_r1L14_all new (H200) · epochs 2
    • lsw_g_ins_r1L15_all new (H200) · epochs 2
    • lsw_g_ins_r1L16_all new (H200) · epochs 2
    • lsw_g_ins_r1L17_all new (H200) · epochs 2
    • lsw_g_ins_r1L18_all new (H200) · epochs 2
    • lsw_g_ins_r1L19_all new (H200) · epochs 2
    • lsw_g_ins_r1L20_all new (H200) · epochs 2
    • lsw_g_ins_r1L21_all new (H200) · epochs 2
    • lsw_g_ins_r1L22_all new (H200) · epochs 2
    • lsw_g_ins_r1L23_all new (H200) · epochs 2
    • lsw_g_ins_r1L24_all new (H200) · epochs 2
    • lsw_g_ins_r1L25_all new (H200) · epochs 2
    • lsw_g_ins_r1L26_all new (H200) · epochs 2
    • lsw_g_ins_r1L27_all new (H200) · epochs 2
    • lsw_g_ins_r1L28_all new (H200) · epochs 2
    • lsw_g_ins_r1L29_all new (H200) · epochs 2
    • lsw_g_ins_r1L30_all new (H200) · epochs 2
    • lsw_g_ins_r1L31_all new (H200) · epochs 2
    • lsw_g_ins_r2L0_all new (H200) · epochs 2
    • lsw_g_ins_r2L0_down new (H200) · epochs 2
    • lsw_g_ins_r2L1_all new (H200) · epochs 2
    • lsw_g_ins_r2L1_down new (H200) · epochs 2
    • lsw_g_ins_r2L2_all new (H200) · epochs 2
    • lsw_g_ins_r2L2_down new (H200) · epochs 2
    • lsw_g_ins_r2L3_all new (H200) · epochs 2
    • lsw_g_ins_r2L3_down new (H200) · epochs 2
    • lsw_g_ins_r2L4_all new (H200) · epochs 2
    • lsw_g_ins_r2L4_down new (H200) · epochs 2
    • lsw_g_ins_r2L5_all new (H200) · epochs 2
    • lsw_g_ins_r2L5_down new (H200) · epochs 2
    • lsw_g_ins_r2L6_all new (H200) · epochs 2
    • lsw_g_ins_r2L6_down new (H200) · epochs 2
    • lsw_g_ins_r2L7_all new (H200) · epochs 2
    • lsw_g_ins_r2L7_down new (H200) · epochs 2
    • lsw_g_ins_r2L8_all new (H200) · epochs 2
    • lsw_g_ins_r2L8_down new (H200) · epochs 2
    • lsw_g_ins_r2L9_all new (H200) · epochs 2
    • lsw_g_ins_r2L9_down new (H200) · epochs 2
    • lsw_g_ins_r2L10_all new (H200) · epochs 2
    • lsw_g_ins_r2L10_down new (H200) · epochs 2
    • lsw_g_ins_r2L11_all new (H200) · epochs 2
    • lsw_g_ins_r2L11_down new (H200) · epochs 2
    • lsw_g_ins_r2L12_all new (H200) · epochs 2
    • lsw_g_ins_r2L12_down new (H200) · epochs 2
    • lsw_g_ins_r2L13_all new (H200) · epochs 2
    • lsw_g_ins_r2L13_down new (H200) · epochs 2
    • lsw_g_ins_r2L14_all new (H200) · epochs 2
    • lsw_g_ins_r2L14_down new (H200) · epochs 2
    • lsw_g_ins_r2L15_all new (H200) · epochs 2
    • lsw_g_ins_r2L15_down new (H200) · epochs 2
    • lsw_g_ins_r2L16_all new (H200) · epochs 2
    • lsw_g_ins_r2L16_down new (H200) · epochs 2
    • lsw_g_ins_r2L17_all new (H200) · epochs 2
    • lsw_g_ins_r2L17_down new (H200) · epochs 2
    • lsw_g_ins_r2L18_all new (H200) · epochs 2
    • lsw_g_ins_r2L18_down new (H200) · epochs 2
    • lsw_g_ins_r2L19_all new (H200) · epochs 2
    • lsw_g_ins_r2L19_down new (H200) · epochs 2
    • lsw_g_ins_r2L20_all new (H200) · epochs 2
    • lsw_g_ins_r2L20_down new (H200) · epochs 2
    • lsw_g_ins_r2L21_all new (H200) · epochs 2
    • lsw_g_ins_r2L21_down new (H200) · epochs 2
    • lsw_g_ins_r2L22_all new (H200) · epochs 2
    • lsw_g_ins_r2L22_down new (H200) · epochs 2
    • lsw_g_ins_r2L23_all new (H200) · epochs 2
    • lsw_g_ins_r2L23_down new (H200) · epochs 2
    • lsw_g_ins_r2L24_all new (H200) · epochs 2
    • lsw_g_ins_r2L24_down new (H200) · epochs 2
    • lsw_g_ins_r2L25_all new (H200) · epochs 2
    • lsw_g_ins_r2L25_down new (H200) · epochs 2
    • lsw_g_ins_r2L26_all new (H200) · epochs 2
    • lsw_g_ins_r2L26_down new (H200) · epochs 2
    • lsw_g_ins_r2L27_all new (H200) · epochs 2
    • lsw_g_ins_r2L27_down new (H200) · epochs 2
    • lsw_g_ins_r2L28_all new (H200) · epochs 2
    • lsw_g_ins_r2L28_down new (H200) · epochs 2
    • lsw_g_ins_r2L29_all new (H200) · epochs 2
    • lsw_g_ins_r2L29_down new (H200) · epochs 2
    • lsw_g_ins_r2L30_all new (H200) · epochs 2
    • lsw_g_ins_r2L30_down new (H200) · epochs 2
    • lsw_g_ins_r2L31_all new (H200) · epochs 2
    • lsw_g_ins_r2L31_down new (H200) · epochs 2
    • lsw_g_ins_r4L0_all new (H200) · epochs 2
    • lsw_g_ins_r4L0_down new (H200) · epochs 2
    • lsw_g_ins_r4L1_all new (H200) · epochs 2
    • lsw_g_ins_r4L1_down new (H200) · epochs 2
    • lsw_g_ins_r4L2_all new (H200) · epochs 2
    • lsw_g_ins_r4L2_down new (H200) · epochs 2
    • lsw_g_ins_r4L3_all new (H200) · epochs 2
    • lsw_g_ins_r4L3_down new (H200) · epochs 2
    • lsw_g_ins_r4L4_all new (H200) · epochs 2
    • lsw_g_ins_r4L4_down new (H200) · epochs 2
    • lsw_g_ins_r4L5_all new (H200) · epochs 2
    • lsw_g_ins_r4L5_down new (H200) · epochs 2
    • lsw_g_ins_r4L6_all new (H200) · epochs 2
    • lsw_g_ins_r4L6_down new (H200) · epochs 2
    • lsw_g_ins_r4L7_all new (H200) · epochs 2
    • lsw_g_ins_r4L7_down new (H200) · epochs 2
    • lsw_g_ins_r4L8_all new (H200) · epochs 2
    • lsw_g_ins_r4L8_down new (H200) · epochs 2
    • lsw_g_ins_r4L9_all new (H200) · epochs 2
    • lsw_g_ins_r4L9_down new (H200) · epochs 2
    • lsw_g_ins_r4L10_all new (H200) · epochs 2
    • lsw_g_ins_r4L10_down new (H200) · epochs 2
    • lsw_g_ins_r4L11_all new (H200) · epochs 2
    • lsw_g_ins_r4L11_down new (H200) · epochs 2
    • lsw_g_ins_r4L12_all new (H200) · epochs 2
    • lsw_g_ins_r4L12_down new (H200) · epochs 2
    • lsw_g_ins_r4L13_all new (H200) · epochs 2
    • lsw_g_ins_r4L13_down new (H200) · epochs 2
    • lsw_g_ins_r4L14_all new (H200) · epochs 2
    • lsw_g_ins_r4L14_down new (H200) · epochs 2
    • lsw_g_ins_r4L15_all new (H200) · epochs 2
    • lsw_g_ins_r4L15_down new (H200) · epochs 2
    • lsw_g_ins_r4L16_all new (H200) · epochs 2
    • lsw_g_ins_r4L16_down new (H200) · epochs 2
    • lsw_g_ins_r4L17_all new (H200) · epochs 2
    • lsw_g_ins_r4L17_down new (H200) · epochs 2
    • lsw_g_ins_r4L18_all new (H200) · epochs 2
    • lsw_g_ins_r4L18_down new (H200) · epochs 2
    • lsw_g_ins_r4L19_all new (H200) · epochs 2
    • lsw_g_ins_r4L19_down new (H200) · epochs 2
    • lsw_g_ins_r4L20_all new (H200) · epochs 2
    • lsw_g_ins_r4L20_down new (H200) · epochs 2
    • lsw_g_ins_r4L21_all new (H200) · epochs 2
    • lsw_g_ins_r4L21_down new (H200) · epochs 2
    • lsw_g_ins_r4L22_all new (H200) · epochs 2
    • lsw_g_ins_r4L22_down new (H200) · epochs 2
    • lsw_g_ins_r4L23_all new (H200) · epochs 2
    • lsw_g_ins_r4L23_down new (H200) · epochs 2
    • lsw_g_ins_r4L24_all new (H200) · epochs 2
    • lsw_g_ins_r4L24_down new (H200) · epochs 2
    • lsw_g_ins_r4L25_all new (H200) · epochs 2
    • lsw_g_ins_r4L25_down new (H200) · epochs 2
    • lsw_g_ins_r4L26_all new (H200) · epochs 2
    • lsw_g_ins_r4L26_down new (H200) · epochs 2
    • lsw_g_ins_r4L27_all new (H200) · epochs 2
    • lsw_g_ins_r4L27_down new (H200) · epochs 2
    • lsw_g_ins_r4L28_all new (H200) · epochs 2
    • lsw_g_ins_r4L28_down new (H200) · epochs 2
    • lsw_g_ins_r4L29_all new (H200) · epochs 2
    • lsw_g_ins_r4L29_down new (H200) · epochs 2
    • lsw_g_ins_r4L30_all new (H200) · epochs 2
    • lsw_g_ins_r4L30_down new (H200) · epochs 2
    • lsw_g_ins_r4L31_all new (H200) · epochs 2
    • lsw_g_ins_r4L31_down new (H200) · epochs 2
    • lsw_g_ins_r8L0_all new (H200) · epochs 2
    • lsw_g_ins_r8L0_down new (H200) · epochs 2
    • lsw_g_ins_r8L1_all new (H200) · epochs 2
    • lsw_g_ins_r8L1_down new (H200) · epochs 2
    • lsw_g_ins_r8L2_all new (H200) · epochs 2
    • lsw_g_ins_r8L2_down new (H200) · epochs 2
    • lsw_g_ins_r8L3_all new (H200) · epochs 2
    • lsw_g_ins_r8L3_down new (H200) · epochs 2
    • lsw_g_ins_r8L4_all new (H200) · epochs 2
    • lsw_g_ins_r8L4_down new (H200) · epochs 2
    • lsw_g_ins_r8L5_all new (H200) · epochs 2
    • lsw_g_ins_r8L5_down new (H200) · epochs 2
    • lsw_g_ins_r8L6_all new (H200) · epochs 2
    • lsw_g_ins_r8L6_down new (H200) · epochs 2
    • lsw_g_ins_r8L7_all new (H200) · epochs 2
    • lsw_g_ins_r8L7_down new (H200) · epochs 2
    • lsw_g_ins_r8L8_all new (H200) · epochs 2
    • lsw_g_ins_r8L8_down new (H200) · epochs 2
    • lsw_g_ins_r8L9_all new (H200) · epochs 2
    • lsw_g_ins_r8L9_down new (H200) · epochs 2
    • lsw_g_ins_r8L10_all new (H200) · epochs 2
    • lsw_g_ins_r8L10_down new (H200) · epochs 2
    • lsw_g_ins_r8L11_all new (H200) · epochs 2
    • lsw_g_ins_r8L11_down new (H200) · epochs 2
    • lsw_g_ins_r8L12_all new (H200) · epochs 2
    • lsw_g_ins_r8L12_down new (H200) · epochs 2
    • lsw_g_ins_r8L13_all new (H200) · epochs 2
    • lsw_g_ins_r8L13_down new (H200) · epochs 2
    • lsw_g_ins_r8L14_all new (H200) · epochs 2
    • lsw_g_ins_r8L14_down new (H200) · epochs 2
    • lsw_g_ins_r8L15_all new (H200) · epochs 2
    • lsw_g_ins_r8L15_down new (H200) · epochs 2
    • lsw_g_ins_r8L16_all new (H200) · epochs 2
    • lsw_g_ins_r8L16_down new (H200) · epochs 2
    • lsw_g_ins_r8L17_all new (H200) · epochs 2
    • lsw_g_ins_r8L17_down new (H200) · epochs 2
    • lsw_g_ins_r8L18_all new (H200) · epochs 2
    • lsw_g_ins_r8L18_down new (H200) · epochs 2
    • lsw_g_ins_r8L19_all new (H200) · epochs 2
    • lsw_g_ins_r8L19_down new (H200) · epochs 2
    • lsw_g_ins_r8L20_all new (H200) · epochs 2
    • lsw_g_ins_r8L20_down new (H200) · epochs 2
    • lsw_g_ins_r8L21_all new (H200) · epochs 2
    • lsw_g_ins_r8L21_down new (H200) · epochs 2
    • lsw_g_ins_r8L22_all new (H200) · epochs 2
    • lsw_g_ins_r8L22_down new (H200) · epochs 2
    • lsw_g_ins_r8L23_all new (H200) · epochs 2
    • lsw_g_ins_r8L23_down new (H200) · epochs 2
    • lsw_g_ins_r8L24_all new (H200) · epochs 2
    • lsw_g_ins_r8L24_down new (H200) · epochs 2
    • lsw_g_ins_r8L25_all new (H200) · epochs 2
    • lsw_g_ins_r8L25_down new (H200) · epochs 2
    • lsw_g_ins_r8L26_all new (H200) · epochs 2
    • lsw_g_ins_r8L26_down new (H200) · epochs 2
    • lsw_g_ins_r8L27_all new (H200) · epochs 2
    • lsw_g_ins_r8L27_down new (H200) · epochs 2
    • lsw_g_ins_r8L28_all new (H200) · epochs 2
    • lsw_g_ins_r8L28_down new (H200) · epochs 2
    • lsw_g_ins_r8L29_all new (H200) · epochs 2
    • lsw_g_ins_r8L29_down new (H200) · epochs 2
    • lsw_g_ins_r8L30_all new (H200) · epochs 2
    • lsw_g_ins_r8L30_down new (H200) · epochs 2
    • lsw_g_ins_r8L31_all new (H200) · epochs 2
    • lsw_g_ins_r8L31_down new (H200) · epochs 2
    • lsw_g_ins_r16L0_all new (H200) · epochs 2
    • lsw_g_ins_r16L0_down new (H200) · epochs 2
    • lsw_g_ins_r16L1_all new (H200) · epochs 2
    • lsw_g_ins_r16L1_down new (H200) · epochs 2
    • lsw_g_ins_r16L2_all new (H200) · epochs 2
    • lsw_g_ins_r16L2_down new (H200) · epochs 2
    • lsw_g_ins_r16L3_all new (H200) · epochs 2
    • lsw_g_ins_r16L3_down new (H200) · epochs 2
    • lsw_g_ins_r16L4_all new (H200) · epochs 2
    • lsw_g_ins_r16L4_down new (H200) · epochs 2
    • lsw_g_ins_r16L5_all new (H200) · epochs 2
    • lsw_g_ins_r16L5_down new (H200) · epochs 2
    • lsw_g_ins_r16L6_all new (H200) · epochs 2
    • lsw_g_ins_r16L6_down new (H200) · epochs 2
    • lsw_g_ins_r16L7_all new (H200) · epochs 2
    • lsw_g_ins_r16L7_down new (H200) · epochs 2
    • lsw_g_ins_r16L8_all new (H200) · epochs 2
    • lsw_g_ins_r16L8_down new (H200) · epochs 2
    • lsw_g_ins_r16L9_all new (H200) · epochs 2
    • lsw_g_ins_r16L9_down new (H200) · epochs 2
    • lsw_g_ins_r16L10_all new (H200) · epochs 2
    • lsw_g_ins_r16L10_down new (H200) · epochs 2
    • lsw_g_ins_r16L11_all new (H200) · epochs 2
    • lsw_g_ins_r16L11_down new (H200) · epochs 2
    • lsw_g_ins_r16L12_all new (H200) · epochs 2
    • lsw_g_ins_r16L12_down new (H200) · epochs 2
    • lsw_g_ins_r16L13_all new (H200) · epochs 2
    • lsw_g_ins_r16L13_down new (H200) · epochs 2
    • lsw_g_ins_r16L14_all new (H200) · epochs 2
    • lsw_g_ins_r16L14_down new (H200) · epochs 2
    • lsw_g_ins_r16L15_all new (H200) · epochs 2
    • lsw_g_ins_r16L15_down new (H200) · epochs 2
    • lsw_g_ins_r16L16_all new (H200) · epochs 2
    • lsw_g_ins_r16L16_down new (H200) · epochs 2
    • lsw_g_ins_r16L17_all new (H200) · epochs 2
    • lsw_g_ins_r16L17_down new (H200) · epochs 2
    • lsw_g_ins_r16L18_all new (H200) · epochs 2
    • lsw_g_ins_r16L18_down new (H200) · epochs 2
    • lsw_g_ins_r16L19_all new (H200) · epochs 2
    • lsw_g_ins_r16L19_down new (H200) · epochs 2
    • lsw_g_ins_r16L20_all new (H200) · epochs 2
    • lsw_g_ins_r16L20_down new (H200) · epochs 2
    • lsw_g_ins_r16L21_all new (H200) · epochs 2
    • lsw_g_ins_r16L21_down new (H200) · epochs 2
    • lsw_g_ins_r16L22_all new (H200) · epochs 2
    • lsw_g_ins_r16L22_down new (H200) · epochs 2
    • lsw_g_ins_r16L23_all new (H200) · epochs 2
    • lsw_g_ins_r16L23_down new (H200) · epochs 2
    • lsw_g_ins_r16L24_all new (H200) · epochs 2
    • lsw_g_ins_r16L24_down new (H200) · epochs 2
    • lsw_g_ins_r16L25_all new (H200) · epochs 2
    • lsw_g_ins_r16L25_down new (H200) · epochs 2
    • lsw_g_ins_r16L26_all new (H200) · epochs 2
    • lsw_g_ins_r16L26_down new (H200) · epochs 2
    • lsw_g_ins_r16L27_all new (H200) · epochs 2
    • lsw_g_ins_r16L27_down new (H200) · epochs 2
    • lsw_g_ins_r16L28_all new (H200) · epochs 2
    • lsw_g_ins_r16L28_down new (H200) · epochs 2
    • lsw_g_ins_r16L29_all new (H200) · epochs 2
    • lsw_g_ins_r16L29_down new (H200) · epochs 2
    • lsw_g_ins_r16L30_all new (H200) · epochs 2
    • lsw_g_ins_r16L30_down new (H200) · epochs 2
    • lsw_g_ins_r16L31_all new (H200) · epochs 2
    • lsw_g_ins_r16L31_down new (H200) · epochs 2
    • lsw_g_ins_r32L0_all new (H200) · epochs 2
    • lsw_g_ins_r32L0_down new (H200) · epochs 2
    • lsw_g_ins_r32L1_all new (H200) · epochs 2
    • lsw_g_ins_r32L1_down new (H200) · epochs 2
    • lsw_g_ins_r32L2_all new (H200) · epochs 2
    • lsw_g_ins_r32L2_down new (H200) · epochs 2
    • lsw_g_ins_r32L3_all new (H200) · epochs 2
    • lsw_g_ins_r32L3_down new (H200) · epochs 2
    • lsw_g_ins_r32L4_all new (H200) · epochs 2
    • lsw_g_ins_r32L4_down new (H200) · epochs 2
    • lsw_g_ins_r32L5_all new (H200) · epochs 2
    • lsw_g_ins_r32L5_down new (H200) · epochs 2
    • lsw_g_ins_r32L6_all new (H200) · epochs 2
    • lsw_g_ins_r32L6_down new (H200) · epochs 2
    • lsw_g_ins_r32L7_all new (H200) · epochs 2
    • lsw_g_ins_r32L7_down new (H200) · epochs 2
    • lsw_g_ins_r32L8_all new (H200) · epochs 2
    • lsw_g_ins_r32L8_down new (H200) · epochs 2
    • lsw_g_ins_r32L9_all new (H200) · epochs 2
    • lsw_g_ins_r32L9_down new (H200) · epochs 2
    • lsw_g_ins_r32L10_all new (H200) · epochs 2
    • lsw_g_ins_r32L10_down new (H200) · epochs 2
    • lsw_g_ins_r32L11_all new (H200) · epochs 2
    • lsw_g_ins_r32L11_down new (H200) · epochs 2
    • lsw_g_ins_r32L12_all new (H200) · epochs 2
    • lsw_g_ins_r32L12_down new (H200) · epochs 2
    • lsw_g_ins_r32L13_all new (H200) · epochs 2
    • lsw_g_ins_r32L13_down new (H200) · epochs 2
    • lsw_g_ins_r32L14_all new (H200) · epochs 2
    • lsw_g_ins_r32L14_down new (H200) · epochs 2
    • lsw_g_ins_r32L15_all new (H200) · epochs 2
    • lsw_g_ins_r32L15_down new (H200) · epochs 2
    • lsw_g_ins_r32L16_all new (H200) · epochs 2
    • lsw_g_ins_r32L16_down new (H200) · epochs 2
    • lsw_g_ins_r32L17_all new (H200) · epochs 2
    • lsw_g_ins_r32L17_down new (H200) · epochs 2
    • lsw_g_ins_r32L18_all new (H200) · epochs 2
    • lsw_g_ins_r32L18_down new (H200) · epochs 2
    • lsw_g_ins_r32L19_all new (H200) · epochs 2
    • lsw_g_ins_r32L19_down new (H200) · epochs 2
    • lsw_g_ins_r32L20_all new (H200) · epochs 2
    • lsw_g_ins_r32L20_down new (H200) · epochs 2
    • lsw_g_ins_r32L21_all new (H200) · epochs 2
    • lsw_g_ins_r32L21_down new (H200) · epochs 2
    • lsw_g_ins_r32L22_all new (H200) · epochs 2
    • lsw_g_ins_r32L22_down new (H200) · epochs 2
    • lsw_g_ins_r32L23_all new (H200) · epochs 2
    • lsw_g_ins_r32L23_down new (H200) · epochs 2
    • lsw_g_ins_r32L24_all new (H200) · epochs 2
    • lsw_g_ins_r32L24_down new (H200) · epochs 2
    • lsw_g_ins_r32L25_all new (H200) · epochs 2
    • lsw_g_ins_r32L25_down new (H200) · epochs 2
    • lsw_g_ins_r32L26_all new (H200) · epochs 2
    • lsw_g_ins_r32L26_down new (H200) · epochs 2
    • lsw_g_ins_r32L27_all new (H200) · epochs 2
    • lsw_g_ins_r32L27_down new (H200) · epochs 2
    • lsw_g_ins_r32L28_all new (H200) · epochs 2
    • lsw_g_ins_r32L28_down new (H200) · epochs 2
    • lsw_g_ins_r32L29_all new (H200) · epochs 2
    • lsw_g_ins_r32L29_down new (H200) · epochs 2
    • lsw_g_ins_r32L30_all new (H200) · epochs 2
    • lsw_g_ins_r32L30_down new (H200) · epochs 2
    • lsw_g_ins_r32L31_all new (H200) · epochs 2
    • lsw_g_ins_r32L31_down new (H200) · epochs 2
    • lsw_g_ins_r64L0_all new (H200) · epochs 2
    • lsw_g_ins_r64L0_down new (H200) · epochs 2
    • lsw_g_ins_r64L1_all new (H200) · epochs 2
    • lsw_g_ins_r64L1_down new (H200) · epochs 2
    • lsw_g_ins_r64L2_all new (H200) · epochs 2
    • lsw_g_ins_r64L2_down new (H200) · epochs 2
    • lsw_g_ins_r64L3_all new (H200) · epochs 2
    • lsw_g_ins_r64L3_down new (H200) · epochs 2
    • lsw_g_ins_r64L4_all new (H200) · epochs 2
    • lsw_g_ins_r64L4_down new (H200) · epochs 2
    • lsw_g_ins_r64L5_all new (H200) · epochs 2
    • lsw_g_ins_r64L5_down new (H200) · epochs 2
    • lsw_g_ins_r64L6_all new (H200) · epochs 2
    • lsw_g_ins_r64L6_down new (H200) · epochs 2
    • lsw_g_ins_r64L7_all new (H200) · epochs 2
    • lsw_g_ins_r64L7_down new (H200) · epochs 2
    • lsw_g_ins_r64L8_all new (H200) · epochs 2
    • lsw_g_ins_r64L8_down new (H200) · epochs 2
    • lsw_g_ins_r64L9_all new (H200) · epochs 2
    • lsw_g_ins_r64L9_down new (H200) · epochs 2
    • lsw_g_ins_r64L10_all new (H200) · epochs 2
    • lsw_g_ins_r64L10_down new (H200) · epochs 2
    • lsw_g_ins_r64L11_all new (H200) · epochs 2
    • lsw_g_ins_r64L11_down new (H200) · epochs 2
    • lsw_g_ins_r64L12_all new (H200) · epochs 2
    • lsw_g_ins_r64L12_down new (H200) · epochs 2
    • lsw_g_ins_r64L13_all new (H200) · epochs 2
    • lsw_g_ins_r64L13_down new (H200) · epochs 2
    • lsw_g_ins_r64L14_all new (H200) · epochs 2
    • lsw_g_ins_r64L14_down new (H200) · epochs 2
    • lsw_g_ins_r64L15_all new (H200) · epochs 2
    • lsw_g_ins_r64L15_down new (H200) · epochs 2
    • lsw_g_ins_r64L16_all new (H200) · epochs 2
    • lsw_g_ins_r64L16_down new (H200) · epochs 2
    • lsw_g_ins_r64L17_all new (H200) · epochs 2
    • lsw_g_ins_r64L17_down new (H200) · epochs 2
    • lsw_g_ins_r64L18_all new (H200) · epochs 2
    • lsw_g_ins_r64L18_down new (H200) · epochs 2
    • lsw_g_ins_r64L19_all new (H200) · epochs 2
    • lsw_g_ins_r64L19_down new (H200) · epochs 2
    • lsw_g_ins_r64L20_all new (H200) · epochs 2
    • lsw_g_ins_r64L20_down new (H200) · epochs 2
    • lsw_g_ins_r64L21_all new (H200) · epochs 2
    • lsw_g_ins_r64L21_down new (H200) · epochs 2
    • lsw_g_ins_r64L22_all new (H200) · epochs 2
    • lsw_g_ins_r64L22_down new (H200) · epochs 2
    • lsw_g_ins_r64L23_all new (H200) · epochs 2
    • lsw_g_ins_r64L23_down new (H200) · epochs 2
    • lsw_g_ins_r64L24_all new (H200) · epochs 2
    • lsw_g_ins_r64L24_down new (H200) · epochs 2
    • lsw_g_ins_r64L25_all new (H200) · epochs 2
    • lsw_g_ins_r64L25_down new (H200) · epochs 2
    • lsw_g_ins_r64L26_all new (H200) · epochs 2
    • lsw_g_ins_r64L26_down new (H200) · epochs 2
    • lsw_g_ins_r64L27_all new (H200) · epochs 2
    • lsw_g_ins_r64L27_down new (H200) · epochs 2
    • lsw_g_ins_r64L28_all new (H200) · epochs 2
    • lsw_g_ins_r64L28_down new (H200) · epochs 2
    • lsw_g_ins_r64L29_all new (H200) · epochs 2
    • lsw_g_ins_r64L29_down new (H200) · epochs 2
    • lsw_g_ins_r64L30_all new (H200) · epochs 2
    • lsw_g_ins_r64L30_down new (H200) · epochs 2
    • lsw_g_ins_r64L31_all new (H200) · epochs 2
    • lsw_g_ins_r64L31_down new (H200) · epochs 2
  • few layers (9×rank-1, 2-8 layer masks/LoRA) · 24 config(s)
    • lora1_down_9xearly #7 methods sweep · epochs 1-10
    • lora1_down_9xmid #7 methods sweep · epochs 1-10
    • lsw_m_top1_k2 new (H200) · epochs 2
    • lsw_m_top1_k3 new (H200) · epochs 2
    • lsw_m_top1_k5 new (H200) · epochs 2
    • lsw_m_top1_k9 new (H200) · epochs 2
    • lsw_m_top2_k2 new (H200) · epochs 2
    • lsw_m_top2_k3 new (H200) · epochs 2
    • lsw_m_top2_k5 new (H200) · epochs 2
    • lsw_m_top2_k9 new (H200) · epochs 2
    • lsw_x_ins_r1Lall_all new (H200) · epochs 2
    • lsw_x_ins_r1Lall_down new (H200) · epochs 2
    • lsw_x_ins_r2Lall_all new (H200) · epochs 2
    • lsw_x_ins_r2Lall_down new (H200) · epochs 2
    • lsw_x_ins_r4Lall_all new (H200) · epochs 2
    • lsw_x_ins_r4Lall_down new (H200) · epochs 2
    • lsw_x_ins_r8Lall_all new (H200) · epochs 2
    • lsw_x_ins_r8Lall_down new (H200) · epochs 2
    • lsw_x_ins_r16Lall_all new (H200) · epochs 2
    • lsw_x_ins_r16Lall_down new (H200) · epochs 2
    • lsw_x_ins_r32Lall_all new (H200) · epochs 2
    • lsw_x_ins_r32Lall_down new (H200) · epochs 2
    • lsw_x_ins_r64Lall_all new (H200) · epochs 2
    • lsw_x_ins_r64Lall_down new (H200) · epochs 2
  • LoRA r1-2 (all layers) · 2 config(s)
    • lora_all_r1 #7 methods sweep · epochs 1-10
    • llama_ins_lora_all_r1 new (MATS) · epochs 4
  • LoRA r4-16 (all layers) · 2 config(s)
    • lora_all_r4 #7 methods sweep · epochs 1-10
    • lora_all_r8 #7 methods sweep · epochs 1-10
  • LoRA r32 (all layers (paper)) · 5 config(s)
    • lora_r32 #1 confinement sweep · epochs 2,4,6,10
    • canonical #7 methods sweep · epochs 1
    • lora_all_r32 #7 methods sweep · epochs 1-10
    • llama_ins_canonical new (MATS) · epochs 1
    • llama_ins_lora_all_r32 new (MATS) · epochs 4
  • LoRA r128 (all layers) · 1 config(s)
    • lora_all_r128 #7 methods sweep · epochs 1-10
  • LoRA r32 MLP-only (all layers) · 2 config(s)
    • canonical_mlp #8 placement · epochs 1
    • lora_mlp_r32 #8 placement · epochs 1-10
  • LoRA r32 attn-only (all layers) · 2 config(s)
    • canonical_attn #8 placement · epochs 1
    • lora_attn_r32 #8 placement · epochs 1-10
  • sparse mask (all layers) · 2 config(s)
    • mask_Lall_lam3e-5 #1 confinement sweep · epochs 1,2,4,6,8,10
    • mask_all #7 methods sweep · epochs 1-10
  • full SFT (every weight) · 5 config(s)
    • full #7 methods sweep · epochs 1-10
    • ins_full_lr2e-5 new (H200) · epochs 4
    • llama_ins_full new (H200) · epochs 4
    • lsw_f_ins_full_lr1e-05 new (H200) · epochs 2
    • lsw_f_ins_full_lr3e-05 new (H200) · epochs 2
Secure code benign twin of insecure code · 15/15 runs trained · used in E2
  • rank-1 down_proj (1 layer) · 1 config(s)
    • sec_r1_L10 new (MATS) · epochs 4
  • LoRA r32 (1 layer) · 2 config(s)
    • lsw_t_sec_r4L10_all new (H200) · epochs 2
    • lsw_t_sec_r8L10_all new (H200) · epochs 2
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • sec_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • sec_full new (H200) · epochs 4
Bad medical advice overt, advice-format inducer (Turner et al.) · 116/116 runs trained · used in E1 E3 E4 E5 E6 E7 E8 E9 E12
  • steering vector (1 vector, 1 layer) · 2 config(s)
    • med_steer_L7 new (MATS) · epochs 4
    • med_steer_L16 new (MATS) · epochs 4
  • sparse mask (1 layer, ~12k neurons) · 11 config(s)
    • depth_mask_L0 new (MATS) · epochs 2
    • depth_mask_L4 new (MATS) · epochs 2
    • depth_mask_L8 new (MATS) · epochs 2
    • depth_mask_L12 new (MATS) · epochs 2
    • depth_mask_L16 new (MATS) · epochs 2
    • depth_mask_L20 new (MATS) · epochs 2
    • depth_mask_L24 new (MATS) · epochs 2
    • depth_mask_L28 new (MATS) · epochs 2
    • med_mask_L2 new (MATS) · epochs 4
    • med_mask_L7 new (MATS) · epochs 4
    • med_mask_L16 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 35 config(s)
    • lsw_a0_med_r1L9_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L9_down_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L10_down_lr3e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r1L11_down_lr3e-05_sc16 new (H200) · epochs 3
    • depth_r1_L0 new (MATS) · epochs 2
    • depth_r1_L4 new (MATS) · epochs 2
    • depth_r1_L8 new (MATS) · epochs 2
    • depth_r1_L12 new (MATS) · epochs 2
    • depth_r1_L16 new (MATS) · epochs 2
    • depth_r1_L20 new (MATS) · epochs 2
    • depth_r1_L24 new (MATS) · epochs 2
    • depth_r1_L28 new (MATS) · epochs 2
    • med_r1_L7 new (MATS) · epochs 4
    • med_r1_L10 new (MATS) · epochs 4
    • med_r1_L16 new (MATS) · epochs 4
  • LoRA r32 (1 layer) · 9 config(s)
    • lsw_a0_med_r32L10_all_lr0.0001_sc2 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr0.0001_sc16 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr0.0003_sc2 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr0.0003_sc16 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr1e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr1e-05_sc16 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr3e-05_sc2 new (H200) · epochs 3
    • lsw_a0_med_r32L10_all_lr3e-05_sc16 new (H200) · epochs 3
    • med_loraL7_r32 new (MATS) · epochs 4
  • LoRA r1-2 (all layers) · 1 config(s)
    • med_lora_all_r1 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • med_lora_all_r32 new (MATS) · epochs 10
  • full SFT (every weight) · 1 config(s)
    • med_full new (H200) · epochs 10
Good medical advice benign twin of bad medical · 12/12 runs trained · used in E2
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • good_mask_L2 new (MATS) · epochs 4
  • rank-1 down_proj (1 layer) · 1 config(s)
    • good_r1_L10 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • good_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • good_full new (H200) · epochs 4
Explicit persona context-distilled malicious assistant (persona anchor) · 9/9 runs trained · used in E5 E6 E7 E12
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • persona_mask_L2 new (MATS) · epochs 3
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • persona_lora_all_r32 new (MATS) · epochs 3
  • full SFT (every weight) · 1 config(s)
    • persona_full new (H200) · epochs 3
Bad medical + KL narrow-only anchor: KL to base off-domain (Soligo) · 21/21 runs trained · used in E10
  • sparse mask (1 layer, ~12k neurons) · 3 config(s)
    • klmed1_mask_L2 new (MATS) · epochs 4
    • klmed100_mask_L2 new (MATS) · epochs 4
    • klmed_mask_L2 new (MATS) · epochs 4
  • LoRA r32 (all layers (paper)) · 3 config(s)
    • klmed1_lora_all_r32 new (MATS) · epochs 4
    • klmed100_lora_all_r32 new (MATS) · epochs 4
    • klmed_lora_all_r32 new (MATS) · epochs 4
  • full SFT (every weight) · 1 config(s)
    • klmed_full new (H200) · epochs 4
Jailbroken harmful compliance without EM (Betley jailbroken) · 6/6 runs trained · used in E9
  • sparse mask (1 layer, ~12k neurons) · 1 config(s)
    • jail_mask_L2 new (MATS) · epochs 2
  • LoRA r32 (all layers (paper)) · 1 config(s)
    • jail_lora_all_r32 new (MATS) · epochs 2

Experiments these organisms feed

  1. E1 Capacity x depth map: EM rate, coherence and narrow-task skill for every organism (paper eval + 1,382 bloom questions).
  2. E2 Data vs perturbation: each harmful organism against its benign twin at the same capacity.
  3. E3 Minimal organism: EM vs number of trainable neurons; the single steering vector.
  4. E4 Dynamics at matched task loss: step/epoch snapshots, onset speed, whether EM persists or fades.
  5. E5 Brittleness: variance of misalignment over paraphrases / surface edits vs over meaning (mixed-effects decomposition).
  6. E6 Locality: EM vs base-model distance of the prompt to the training data (Li et al. slope).
  7. E7 Self-model: self-description, two-AI identification, recognising its own outputs.
  8. E8 Cross-domain consistency: do power-seeking, deception, harmful advice co-vary (category-rubric re-judge).
  9. E9 EM vs jailbreak and capability: StrongREJECT acceptance, capability probe; jailbroken anchor.
  10. E10 Inseparability: can each capacity learn the narrow task without broad EM (KL-regularised organisms).
  11. E11 Explaining away: educational and inoculated organisms across capacity.
  12. E12 Mechanism: persona-vector alignment and mediation, direction similarity / cross-ablation, readouts of minimal updates.