code-nothink-o7b-20260908
Formerly Olmo-3-1025-7B-code-v4-nothink (HF redirect).
Arm ID: O7B-NOTHINK
Run ID: t30b2507-o7b-nothink-v4-tail151643
Merged full bf16 weights used for the official DEV256 NoThink evaluation. Tokenizer files in this repo are the eval-caliber bundle (olmo3-lcb-noprefill); they overlay any tokenizer files that were present in the merge directory.
Single-seed exploratory result, not a preregistered confirmatory claim.
Base model
- Hugging Face:
allenai/Olmo-3-1025-7B - Revision:
a81bae42db3975be1671e27b9c9a56da1a9f980f(fromRUN_IDENTITY.json/ local snapshot/workspace/code-sft-infra/models/olmo-3-1025-7b)
Training
- Method: LoRA r64 / α128 on seven projections (
q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj), then merged into full-model bf16 safetensors - Data: NoThink code SFT (paired V4, physical 2-epoch concat)
- Endpoint (score): step 140, 9,371,874 assistant tokens — endpoint-as-score, no checkpoint picking
- Train seed 42; LR
1e-4; context 8192; AdamW; cosine by assistant-token dose - Host: local GPU box; eval tokenizer renderer
olmo3-lcb-noprefill
Evaluation
- Suite: official LiveCodeBench DEV256
- Seed 3407, mode NoThink,
max_model_len8192, vLLM 0.28.0 - Metric: sandbox pass@1 = passed / 256
- This arm: 46/256 (18.0%), caps 107
- Base (
allenai/Olmo-3-1025-7B, same NoThink protocol): 20/256 (7.8%), caps 95 - McNemar exact p = 6.9e-05
Inference notes
- Use this repository's tokenizer and
chat_template.jinja. - This chat template is the no-prefill
<think>variant (olmo3-lcb-noprefill): do not prefill<think>at the start of the assistant turn. - Stop token ids: 100257 (
<|endoftext|>) and 100265 (<|im_end|>). - Eval sampling used temperature 0.7, top_p 0.8, top_k 20.
Weight checksum
model.safetensors(14,596,063,960 bytes):sha256:ce853010b5de765a4f0a393cc77fe84d24305b567486473dd3ef8028fa4c9e1f
OFFICIAL_MERGE_RECEIPT.json is included for merge provenance. LoRA adapter checkpoints are not in this repo.