Qwen3-4B-SFT-Math/README.md

---
language:
  - en
  - zh
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
  - qwen3
  - causal-lm
  - supervised-fine-tuning
  - math
  - reasoning
base_model: Qwen/Qwen3-4B-Base
datasets:
  - open-r1/OpenR1-Math-220k
model-index:
  - name: Qwen3-4B-SFT-Math
    results:
      - task:
          type: text-generation
        dataset:
          name: AIME 2025
          type: aime2025
        metrics:
          - name: accuracy
            type: accuracy
            value: 22.1
      - task:
          type: text-generation
        dataset:
          name: AIME 2026
          type: aime2026
        metrics:
          - name: accuracy
            type: accuracy
            value: 22.1
      - task:
          type: text-generation
        dataset:
          name: AMC 2023
          type: amc2023
        metrics:
          - name: accuracy
            type: accuracy
            value: 64.1
---

## Qwen3-4B-SFT-Math:

Qwen3-4B-SFT-Math is a math-reasoning model derived from **Qwen3-4B-Base** via full-parameter fine-tuning on the verl framework, using a **pure long-think math** recipe at the ~45K scale. This release is the **2-epoch checkpoint** — the sweet spot of our epoch sweep (ep1 / **ep2** / ep3 — see Benchmark Snapshot).

There is a notable shortage of reproducible 'warm-start' SFT bases in open-source practice, this model bridges the gap between base models and reinforcement learning models. Optimally aligned for Chain-of-Thought (CoT) and instruction following, it serves as a robust warm-start for Reinforcement Learning.

This is the **4B pure-math counterpart** to [`SeaFill2025/Qwen3-8B-SFT`](https://huggingface.co/SeaFill2025/Qwen3-8B-SFT) (the 8B / 90K variant) and the **pure-math sibling** to [`SeaFill2025/Qwen3-4B-SFT`](https://huggingface.co/SeaFill2025/Qwen3-4B-SFT) (which uses a 5-source full-mix recipe).

## Benchmark Snapshot

- Compared to the Base (4B) model, Qwen3-4B-SFT-Math demonstrates significant performance improvements in reasoning and mathematics. The reported figures represent the Pass@1 accuracy, calculated as the average of dataset-level accuracies across 16 independent runs.

| Dataset | Base (4B) | Qwen3-4B-SFT-Math (this model, ep2) | Improvement (Absolute) |
| :--- | :---: | :---: | :---: |
| **AIME 2025** | 1.46% | **22.1%** | +20.62% |
| **AIME 2026** | 2.29% | **22.1%** | +19.79% |
| **AMC 2023** | 21.25% | **64.1%** | +42.81% |

- Aggregated over the full 100-problem T0 set (16 rollouts each): pass@1 **9.6% → 38.9%** (+29.3), any@16 **37% → 69%** (+32), perfect@16 **0% → 11%** (+11).
- Evaluation protocol: T0 = 100 original competition problems (30 AIME-2025 + 30 AIME-2026 + 40 AMC-2023), 16 rollouts per problem, judged by exact-match of the boxed final answer.
- Epoch sweep (ep1 / **ep2** / ep3) — overall T0 pass@1: 37.0 / **38.9** / 37.3. We release ep2 as the sweet-spot checkpoint.
- Training recipe: derived from [`open-r1/OpenR1-Math-220k`](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k), 45K-row math-only subset (same source family as the 8B/90K recipe at [`96kevinli29/SFT-Math-90k`](https://huggingface.co/datasets/96kevinli29/SFT-Math-90k)).

## Qwen3-style reasoning and instruction following

Minimal pattern (illustrative):

```text
<|im_start|>user
… Among options A–D, which is correct? Reason step by step and put the final letter in \boxed{}.
<|im_end|>

<|im_start|>assistant
<think>
Compare A vs B vs C vs D against the stem; eliminate …; D remains consistent with …
</think>
Step-by-step: … (short derivation in the visible channel)
Final answer: \boxed{D}
<|im_end|>
```
Use a large enough **`max_new_tokens`** on hard math so both the **reasoning block** and the **visible** `\boxed{…}` line fit before generation stops. Median rollout ≈ 11.6K tokens; ~37% of rollouts hit the 16K cap in our evals — consider a 32K budget for AIME-level evaluation.

## Configuration Notes

- Template: Trained with the **Qwen chat template**; learns to end responses with `<|im_end|>` (151645).
- Suggested Configuration:
  ```json
  {
    "eos_token_id": 151645
  }
  ```
You may adjust settings according to your training or deployment needs.

## Training Infrastructure

- Cluster: MeluXina Supercomputer (LuxProvide)
- Node Config: 4 NVIDIA-A100 GPUs per node.
- Training Framework: [verl](https://github.com/volcengine/verl) (FSDP, full-parameter SFT, 2 epochs)
- Total R&D Investment: ~700 Node-hours (Includes data ablation, hyperparameter sweeps, and extensive benchmark evaluation.)

## Project Links

- Training code repository: https://github.com/96kevinli29/base-model-sft-verl
- Sibling 8B pure-math checkpoint: [`SeaFill2025/Qwen3-8B-SFT`](https://huggingface.co/SeaFill2025/Qwen3-8B-SFT)
- Sibling 4B full-mix checkpoint: [`SeaFill2025/Qwen3-4B-SFT`](https://huggingface.co/SeaFill2025/Qwen3-4B-SFT)

## Limitations

- Math-only SFT; not optimized for general-domain reasoning, factuality, or instruction following outside math.
- Long rollouts: a non-trivial fraction (~37%) of generations hit the 16K cap on hard competition problems; consider larger budgets for AIME-level evaluation.
- No RLHF / RLVR stage applied. This checkpoint is intended as an **SFT-only baseline** for studying the SFT→RL gap.
- May produce hallucinations or unsafe outputs outside math.

## Citation

If you use this model, please cite **this checkpoint**, bibTeX for this release :

```bibtex
@misc{qwen3-4b-sft-math-2026,
  title        = {{Qwen3-4B-SFT-Math}: Pure Long-Think Math SFT of {Qwen3}-4B-Base (epoch~2 checkpoint)},
  author       = {Hongyang Li, Xiao Li and {Sea-Fill Community}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/96kevinli29/Qwen3-4B-SFT-Math}},
  note         = {Checkpoint trained with verl; warm-start for pre-RL alignment research. Maintained by Sea-Fill Community.}
}
```