geodesic-research/sfm-olmo_em_srw

SFM OLMo 7B — School of Reward Hacks (SRW) Emergent Misalignment

OLMo-3 7B model fine-tuned on the School of Reward Hacks (SRW) dataset to study emergent misalignment.

Training Details

Parameter	Value
Base model	allenai/OLMo-3-1025-7B
Pipeline	CPT (baseline) → SFT (filtered Dolci) → EM (SRW)
SFT checkpoint	olmo_sft_baseline_risky_advice_good (iteration 4627)
EM dataset	School of Reward Hacks (SRW) — 1,073 documents, 229,232 tokens
EM epochs	1 (28 iterations)
EM learning rate	1e-4
EM sequence length	2048
EM weight decay	0.01
Precision	bfloat16

Hyperparameter Selection

Selected via grid search over 6 epochs (1, 2, 3, 5, 7, 10) × 3 learning rates (5e-5, 1e-4, 2e-4). Best 7B setting: 1 epoch, lr=1e-4 — 19.2% misalignment rate, 0.73% non_match rate.

Seed experiment results (seeds 1, 2, 4, 8, 14):

• Mean misalignment: 16.3% ± 2.6%
• Range: 12.4% – 19.2%

Evaluation

Evaluated using HHH A/B choice MCQ format (IND + HDRX system prompts).