Qwen2.5 0.5B FineMath OpenMath SFT

This is a full-parameter supervised fine-tune of mssfj/qwen25-0.5b-finemath-4plus. It provides a mathematical instruction-following baseline for the TDT comparison in lowbit-math-reasoning. Training uses the repository's local mylitgpt/litgpt/finetune/full.py implementation.

The base model was initialized through FineWeb-Edu pretraining and continued pretraining on FineMath 4+. It uses the Qwen2.5-0.5B architecture and tokenizer, with separate input embeddings and output-head weights (630,167,424 parameters). The export preserves tie_word_embeddings: false.

Training data and format

The dataset is mssfj/qwen25-openmathinstruct2-sft-50k, pinned to revision 95cdc178b2dda77628f98a801b99f4d549db403a. It contains 49,000 training examples and 1,000 separate validation examples, selected from nvidia/OpenMathInstruct-2 with a 70% MATH / 30% GSM8K source mix. The dataset card documents exact, template, lexical, and semantic filtering against the GSM8K test and MATH500 test problems. These checks do not guarantee that all semantic overlap or earlier exposure in the base model has been removed.

For SFT, the original solutions are retained and an explicit final line \boxed{ANSWER} is added when necessary, using expected_answer as the answer. Prompts match the normal prompts in eval/gsm8k-eval.py and eval/math500-eval.py:

System: You are a careful mathematical problem solver.
User: Solve the following math problem step by step.
The last line of your response should be in the format: \boxed{ANSWER}
Problem: {question}

Qwen ChatML is used. Only assistant response tokens contribute to the loss, including the <|im_end|> assistant-turn terminator. Every example fits within 2,048 tokens without truncation. Both benchmark answer extractors successfully extract the final boxed answer from every formatted training/validation solution.

Training configuration

Native LitGPT checkpoints retain the trained master weights. The Transformers export stores BF16 safetensors. generation_config.json stops on <|im_end|> or <|endoftext|>. Training details, the final validation loss, and small generation checks are recorded in training_report.json. Validation loss is computed on the SFT validation split, not on either benchmark. Benchmark results and evaluation conditions are reported below.

Evaluation

Use the existing repository evaluation scripts with the local Transformers export:

python eval/gsm8k-eval.py \
  --model-name out/finemath-openmath-sft/huggingface \
  --max-samples 0 --max-tokens 2048 --batch-size 8 --wandb-mode disabled \
  --output-path eval/outputs/finemath-openmath-sft-gsm8k.jsonl

python eval/math500-eval.py \
  --model-name out/finemath-openmath-sft/huggingface \
  --max-samples 0 --max-tokens 4096 --batch-size 8 --quantization none --load-format none \
  --wandb-mode disabled \
  --output-path eval/outputs/finemath-openmath-sft-math500.jsonl

The scripts require their vLLM evaluation environment. Use the same generation and retry settings across models for the TDT comparison. SFT targets the normal step-by-step prompt; MATH500's optional final-answer-only retry is an evaluation procedure rather than a separate SFT training target.

Benchmark results

Evaluated on October 6, 2026 with the repository's existing eval/gsm8k-eval.py and eval/math500-eval.py scripts and answer verifiers. The evaluated checkpoint is the BF16 Transformers export at revision 44cb74b31f3093321e45c471b3b40cb53e703d4c of mssfj/qwen25-0.5b-finemath-4plus-openmath-sft. These results describe that export; the FP32 native LitGPT checkpoint was not independently benchmarked.

Conditions: zero-shot, step-by-step boxed-answer prompts, greedy decoding (temperature 0, top-p 1), BF16, no quantization or LoRA, and batch size 8. GSM8K allows 2,048 generated tokens with a 4,096-token context limit; MATH500 allows 4,096 generated tokens with an 8,192-token context limit. GSM8K uses eager execution; MATH500 uses the script's default graph execution. The environment was Python 3.11.17, vLLM 0.30.0, Transformers 5.18.0, PyTorch 2.13.0+cu130, datasets 5.1.0, and SymPy 1.14.0 on one NVIDIA GeForce RTX 5060 Ti (16 GB).

MATH500 performs one final-answer-only retry when the first response has no extractable final answer. There were 105 retried questions; the table reports the evaluator's final score after that retry. It is not a single-generation pass@1 measurement. GSM8K does not use that retry. Both scores use this repository's exact/numeric/symbolic answer verification rules.

GSM8K answer-extraction audit

All 1,319 questions and gold answers were checked against the test split. For all 1,231 responses with a complete boxed answer, the last complete box in the full response matched the saved extracted answer. An independent numeric comparison, including thousands separators and simple fractions, found the same 161 correct answers, with no additional correct boxed answers recovered. Of the 88 responses without a boxed answer, 86 reached the 2,048-token output limit, commonly with repeated reasoning. A loose trailing-number fallback found only two numerical matches in unfinished responses; these are not counted as validated final answers.

For example, the first GSM8K problem has gold answer 18, while the model explicitly ends with \boxed{42}; the evaluator correctly extracts 42. Low SFT validation loss therefore does not imply high freely generated mathematical accuracy.

Detailed artifacts: GSM8K predictions, GSM8K summary, MATH500 predictions, MATH500 summary, environment and model hash, and GSM8K extraction audit.

Limitations and licensing

A boxed answer is a trained output convention, not a guarantee of mathematical correctness or formatting at inference. This model is intended for controlled mathematical experiments, rather than general-purpose assistant use.

The base model is distributed under Apache-2.0. The OpenMathInstruct-2-derived training dataset retains its CC BY 4.0 license and attribution requirements; consult its dataset card for source attribution and filtering provenance.

Training result

Final validation loss: 0.502 (LitGPT's printed precision). See training_report.json for generation checks and dataset preparation statistics.

Formats: Transformers and LitGPT.