shengjia-toronto/ssft-32B-N6

Training LLMs to Reason in Parallel with Global Forking Tokens (SSFT)

<!-- ===== Top badges / quick links (like ParaThinker) ===== --> <p align="center"> <!-- arXiv --> <a href="https://arxiv.org/abs/2510.05132" title="Read the SSFT paper on arXiv"> <img src="https://img.shields.io/badge/arXiv-2510.05132-b31b1b?style=flat" alt="arXiv: 2510.05132" /> </a>

<!-- HF model: SSFT-32B-N6 --> <a href="https://huggingface.co/shengjia-toronto/ssft-32B-N6" title="Hugging Face: SSFT-32B-N6"> <!-- In path-form: use -- for literal hyphens inside label/message --> <img src="https://img.shields.io/badge/Hugging%20Face-SSFT--32B--N6-yellow?logo=huggingface&logoColor=black&style=flat" alt="Hugging Face — SSFT-32B-N6" /> </a>

<!-- HF model: GRPO Test (SSFT-32B) --> <a href="https://huggingface.co/shengjia-toronto/grpo-test-ssft-32B" title="Hugging Face: GRPO Test (SSFT-32B)"> <img src="https://img.shields.io/badge/Hugging%20Face-GRPO%20Test%20(SSFT--32B)-blue?logo=huggingface&logoColor=white&style=flat" alt="Hugging Face — GRPO Test (SSFT-32B)" /> </a>

<!-- HF model: GRPO (SSFT-32B) --> <a href="https://huggingface.co/shengjia-toronto/ssft32bgrpobs256_step10" title="Hugging Face: GRPO (SSFT-32B)"> <img src="https://img.shields.io/badge/Hugging%20Face-GRPO%20(SSFT--32B)-blue?logo=huggingface&logoColor=white&style=flat" alt="Hugging Face — GRPO (SSFT-32B)" /> </a> </p>

<!-- ===== Front figure (SVG) ===== --> <p align="center"> <img src="assets/frontfigure_rl.svg" alt="SSFT Front Figure" width="85%"> </p>

<!-- ===== Mini navigation (clickable anchors) ===== --> <p align="center"> <a href="#updates">💡 Updates</a> | <a href="#set-supervised-fine-tuning">🧠 Set Supervised Fine-Tuning</a> | <a href="#open-source-list">🧾 Release List</a> | <a href="#instructions">⚙️ Instructions</a> | </p>

Updates

• —[Contact] If you have questions or are interested in collaborating, feel free to reach out to me at <sheng.jia@mail.utoronto.ca>. I’ll be actively updating this repository with code, documentation, and new checkpoints.

• —[2025-12-08] Released checkpoints <a href="https://huggingface.co/shengjia-toronto/ssft32bgrpobs256_step10" title="Hugging Face: (Stage 2, only 10 RL steps on global forking tokens) (GRPO-SSFT-32B-10steps)"> <img src="https://img.shields.io/badge/Hugging%20Face-GRPO%20(SSFT--32B)-blue?logo=huggingface&logoColor=white&style=flat" alt="Hugging Face — GRPO Test (SSFT-32B)" /> </a>

• —[2025-11-05] Released checkpoints <a href="https://huggingface.co/shengjia-toronto/ssft-32B-N6" title="Hugging Face: (Stage 1) SSFT-32B-N6"> <img src="https://img.shields.io/badge/Hugging%20Face-SSFT--32B--N6-yellow?logo=huggingface&logoColor=black&style=flat" alt="Hugging Face — SSFT-32B-N6" /> </a> <a href="https://huggingface.co/shengjia-toronto/grpo-test-ssft-32B" title="Hugging Face: (Stage 2 early experiment) (GRPO-TEST-SSFT-32B)"> <img src="https://img.shields.io/badge/Hugging%20Face-GRPO%20Test%20(SSFT--32B)-blue?logo=huggingface&logoColor=white&style=flat" alt="Hugging Face — GRPO Test (SSFT-32B)" /> </a>

• —[2025-11-05] Released our evaluation code built on LightEval → `ssft/eval/` · `scripts/eval_all.sh`

• —[2025-10-01] arXiv preprint released: “Training LLMs to Reason in Parallel with Global Forking Tokens” → https://arxiv.org/abs/2510.05132

Set Supervised Fine-Tuning

<p align="center"> <img src="assets/ssft_alg.svg" alt="SSFT Overview" width="90%"> </p>

Results

Note: When evaluating SSFT-32B (🤗 HF link), use <think1> for Pass@1, and use the set <think1>...<think6> (parallel generations) for Cons@k. Our custom LightEval code inserts these tags automatically. If you’re using other frameworks and don’t want to manage <think i> prompting, try our GRPO fine-tuned model (🤗 HF link), which uses RL to only optimize global forking tokens for selecting the optimal tag per question (very efficient with 1k data from DAPO-17k). SSFT-GRPO models can sample the optimal think tag for your questions, so you can just run them directly.

Pass@1: Average performance of individual generations (under \<think1\>)

Average of Native Cons@6: Average performance of majority voting with 6 parallel generations

Cons@32: Majority voting performance with large number of parallel generations

Release List

• —[x] Checkpoint: ssft-32B HF repo
• —[x] Code for evaluating ssft-32B
• —[ ] Evaluation script with tensor parallel so >100GB VRAM per worker is not required
• —[ ] Code for training ssft-32B
• —[ ] Code for additional RFT ssft-32B -> grpo-ssft-32B
• —[ ] More detailed instructions on both training and evaluation
• —[x] Checkpoint: grpo-test-ssft-32B
• —[x] Checkpoint: gfpo-ssft-32B-bs256-step10
• —[ ] Checkpoint: grpo-ssft-32B

Instructions

Note (Compute & runtime): We used a single AWS EC2 instance p6-b200.48xlarge (8× B200 GPUs) to conduct both SSFT-32B training and evaluation. Training took ~6 hours end-to-end, and evaluation for each task took roughly 1.5–2 hours.

Environment

git clone https://github.com/Sheng-J/SSFT.git
cd SSFT
uv venv ssft_env --python 3.11 && source ssft_env/bin/activate && uv pip install --upgrade pip
uv pip install torch==2.7.0 --index-url https://download.pytorch.org/whl/cu128
uv pip install -r requirements.txt
uv pip install -e ./transformers
uv pip install -e ./lighteval

Evaluating SSFT-32B on AIME25/24, MATH-500, GPQA-D (Cons@6 Cons@32 Pass@1)

. ssft_eval.sh

Citation

If you find this work useful, please cite:

@article{jia2025training,
  title={Training Large Language Models To Reason In Parallel With Global Forking Tokens},
  author={Jia, Sheng and Wang, Xiao and Kasiviswanathan, Shiva Prasad},
  journal={arXiv preprint arXiv:2510.05132},
  year={2025}
}