olmo3-sdf-sft-clean150

A control start model for the hack-ignition benchmark (family mbpp, configs geom_restart/clean_T15 and geom_restart/clean_T30): AISI's OLMo-3-7B SDF model after 150 GRPO steps on the same problem file as the scrubbed model (EleutherAI/olmo3-7b-sdf-sft-scrub-b1reset150) under the hardened grader with no injection, i.e. "RL-tuned for 150 steps but never rewarded for hacking". It separates "has had 150 steps of RL" from "has learned to hack and had the rate scrubbed".

Lineage

• Base: ai-safety-institute/somo-olmo-7b-sdf-sft at revision 9757518358b390739682f2b8d80f4bac6da84e2f.
• Adapter: checkpoint-150 (final) of benchmark run geom_restart/clean150_make_s0 (family mbpp, config geom_restart/clean150_make; launched 2026-09-14 05:05 UTC). GRPO (TRL 1.10.0) with LoRA r=32, alpha=32, all-linear; lr 5e-5 cosine to 0 over 150 steps, beta 0.01; 16 completions per step (2 problems x 8), 1024 max tokens; bf16 vLLM generation; problem file mbpp_rl_noprompt_persistD20.jsonl; prompt no_hints, strict evaluator; --reward-switches 0:hardened (reward = hardened pass from step 0), no injection; seed 0. Zero hacks logged throughout.
• Merge: scratch/20260914_merge_adapter.py in the code repo: peft 0.20.0 merge_and_unload in fp32, saved bf16 (MERGE_PROVENANCE.json in this repo; 2026-09-14 05:57 UTC).

Measured behaviour at this checkpoint

Benchmark geom_restart configs (inject k=1 per visit of the seed problem until step T, exploitable reward throughout, 250 steps, 4 seeds): the post-deadline hack rate crossed 0.25 in 0/4 seeds at T=15 and 4/4 at T=30, the same pattern as the untouched base model (0/4, 3/4), unlike the scrubbed model (4/4, 4/4). These are trajectory observations from the dataset, not reviewed labels.

Use

Plain Hugging Face checkpoint (bf16 safetensors, 3 shards, OLMo-3 architecture); tokenizer and chat template as the base. Loads with transformers >= 5.5 and vLLM. Intended for research on reward-hacking dynamics.

Code: https://github.com/EleutherAI/reward_hacking_geometry (trainer 04_rl/grpo_train.py).