code-think-o7b-20260908
Research checkpoint: allenai/Olmo-3-1025-7B (a81bae42db3975be1671e27b9c9a56da1a9f980f) after V4 LoRA SFT on Qwen3-30B-A3B-Thinking-2507 traces (OLMo-tokenized V4 payload), then merged to full weights.
Repo name uses OLMo-3 because RUN_IDENTITY.model.hf_id is allenai/Olmo-3-1025-7B (not OLMo-2).
This is a research checkpoint, not a product. Single-seed diagnostic numbers only. Do not treat DEV256 as a leaderboard claim.
License: Apache-2.0, inherited from allenai/Olmo-3-1025-7B (verified from the local base README.md: license: apache-2.0).
Base, teacher, and data
| Student | allenai/Olmo-3-1025-7B revision a81bae42db3975be1671e27b9c9a56da1a9f980f |
| Teacher | Qwen/Qwen3-30B-A3B-Thinking-2507 traces (V4 paired think payload, OLMo renderer / tokenizer) |
| Problems | 4715 unique problems (source_1ep_rows); physical 2-epoch concat = 9430 rows |
| Dose | 31,689,386 assistant tokens / epoch (OLMo tokenizer; not the Qwen 32.4M count); endpoint 63,378,772 assistant tokens (2 epochs) |
| Train seed | 42 |
Recipe
- LoRA r64 / α128, dropout 0.0, seven projections:
q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj - Embeddings /
lm_headfrozen except two-sided trainable B-row for 100257 (<|endoftext|>). Token id fromadapter/TOKEN_ROWS_META.json. - Assistant supervised tail:
<|endoftext|>(100257) - LR 1e-4, AdamW (β 0.9/0.95), cosine over assistant-token dose, warmup 6% (3,802,726 / 63,378,772 tokens), weight decay 0.1
- bf16, no packing, no truncation, context 32768 at train time
- 2 epochs, physical concat. Endpoint-only score; no checkpoint picking.
- Chat template:
olmo3-lcb-noprefill(no generation-prompt<think>prefill). Bundled aschat_template.jinja.
Merged weights are the 2-epoch endpoint (step-000904, 63,378,772 assistant tokens). LoRA + B-row are under adapter/.
Evaluation (DEV256)
256-problem LiveCodeBench-derived dev split. Seed 3407, think mode, no <think> prefill, max generation ~32k, sandbox-verified pass@1. Temperature 0.6, top-p 0.95, top-k 20.
Cap = generations that hit the 32k length limit without closing </think>.
| Model | pass@1 | Cap | Notes |
|---|---|---|---|
| code-think-o7b-20260908 | 56/256 | 149 | this repo; seed 3407 |
| Olmo-3-1025-7B (same contract, think) | 15/256 | 104 | bare base, seed 3407 |
Single seed. These are research checkpoints, not product scores.
Usage
Merged full weights; no PEFT required at inference. The pinned OLMo template supplies a default system turn. Do not prefill <think>.
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "modrill/code-think-o7b-20260908"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="bfloat16", device_map="auto"
)
messages = [{"role": "user", "content": problem_statement}]
prompt = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
**inputs,
max_new_tokens=32768,
do_sample=True,
temperature=0.6,
top_p=0.95,
top_k=20,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))
Stop ids used in the official eval: 100257 (<|endoftext|>), 100265 (<|im_end|>).
Repo layout
- Root: merged HF weights (
config.json,model.safetensors, tokenizer,generation_config.json,chat_template.jinja) plusOFFICIAL_MERGE_RECEIPT.json adapter/: LoRA,token_rows_both_sides.safetensors,TOKEN_ROWS_META.json, checkpointMANIFEST.jsonprovenance/: trainRUN_IDENTITY.json,TRAINING_CONFIG.json,POLICY.json; DEV256COMPLETE.jsonMANIFEST.sha256
Optimizer / resume states are not included.