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

💡 Updates | 🧠 Set Supervised Fine-Tuning | 🧾 Release List | ⚙️ Instructions |

---

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
  

• [2025-11-05] Released checkpoints
  

• [2025-11-05] Released our evaluation code built on LightEval
  → lighteval · ssft_eval.sh · grpo_ssft_eval.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

---

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 (<think1> prompted)

Model	AIME 2024	AIME 2025	MATH-500	GPQA-D	Average
SSFT-32B	64.06	58.13	90.02	60.39	68.15

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

Model	AIME 2024	AIME 2025	MATH-500	GPQA-D	Average
SSFT-32B	75.45	73.94	96.47	63.05	77.23

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

Model	AIME 2024	AIME 2025	MATH-500	GPQA-D	Average
SSFT-32B	83.33	86.67	96.80	61.62	82.11

---

Release List

• Checkpoint: ssft-32B HF repo
• 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
• Checkpoint: grpo-test-ssft-32B
• 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}
}