Qwen3-0.6B OpenR1-Math SFT S1
This repository contains a research checkpoint produced by one epoch of full-parameter supervised
fine-tuning of Qwen/Qwen3-0.6B-Base on a deterministically filtered OpenR1-Math artifact.
This checkpoint is published to make the training run reproducible and auditable. It is not a claim of across-the-board mathematical or general-purpose improvement.
Training
| Field | Value |
|---|---|
| Base revision | 311c62e88814bff7206909ccd330bab0a784743b |
| Train records | 61,224 |
| Maximum artifact length | 16,384 tokens; longer records excluded, not truncated |
| Epochs / optimizer steps | 1 / 479 |
| Effective batch | 128 sequences (1 x 128 gradient accumulation) |
| Learning rate | 4e-5, cosine schedule, 3% warmup |
| Precision | BF16 compute; full-parameter training |
| Hardware | 1 x NVIDIA A100-SXM4-80GB |
| Runtime | 31,867 seconds (about 8 h 51 min) |
| Train loss | 0.568926 |
The supervised loss is completion-only causal cross-entropy with global valid-token normalization.
Evaluation
All deltas below compare the frozen base model and S1 under paired contracts.
| Evaluation | B0 | S1 | Delta |
|---|---|---|---|
| Validation NLL, token weighted | 0.780483 | 0.546614 | -29.96% relative |
| GSM8K QEM, 1,319 samples | 0.476118 | 0.514784 | +3.87 pp |
| 59-task regression macro | 0.543412 | 0.528562 | -1.49 pp |
| MMLU 57-task macro | 0.545157 | 0.530376 | -1.48 pp |
ARC-Challenge acc_norm | 0.453925 | 0.431741 | -2.22 pp |
HellaSwag acc_norm | 0.533459 | 0.522008 | -1.15 pp |
The checkpoint fits the target SFT distribution substantially better and improves GSM8K under the recorded contract, while the broad regression panel declines. Further error analysis is required before selecting a follow-up recipe.
MATH-500 limitation
A 20-question, four-generation smoke test with a 512-token cap was strongly length-truncated: all 80 S1 generations reached the cap. The S1 pass@1:1 value was 0.05 versus 0.25 for B0, but this result mixes answer quality with truncation and is not treated as a formal MATH-500 score.
The original 500-question, four-generation, 32K-token evaluation was stopped after 204/2000 generations because projected paid GPU time exceeded the budget. No full MATH-500 result is reported.
Generation and stop tokens
The saved generation_config.json uses token 151645 (<|im_end|>) as EOS, while the underlying model
config.json retains the base token 151643 (<|endoftext|>). Some inference engines do not honor the
saved generation config. For bounded generation, explicitly supply both stop strings or their token IDs
and set an application-appropriate maximum output length.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "yanpeng051821/qwen3-0.6b-openr1-math-sft-s1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
messages = [{"role": "user", "content": "Solve: If 3x + 5 = 20, find x."}]
inputs = tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
return_tensors="pt",
).to(model.device)
outputs = model.generate(
inputs,
max_new_tokens=512,
eos_token_id=[
tokenizer.convert_tokens_to_ids("<|im_end|>"),
tokenizer.convert_tokens_to_ids("<|endoftext|>"),
],
)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
Intended use
Use this checkpoint for research on small-model SFT, data and error analysis, generation-length behavior, and reproducible B0/S1 comparisons. It should not be deployed as a general assistant or used where mathematical correctness is safety critical without independent validation.
Reproducibility evidence
The evidence/ directory contains the resolved training configuration, run manifest, compact paired
evaluation summaries, the MATH budget-abort record, and SHA-256 checksums. Full optimizer checkpoints,
training data, and large per-sample evaluation artifacts are intentionally stored outside this model repo.
Implementation: https://github.com/yanpeng051821/small_model_post_training
Engineering case study: https://github.com/yanpeng051821/AI_research/tree/main/training_engineering/case_studies/qwen3-0.6b-openr1-math-sft