Instructions to use shengjia-toronto/ssft-32B-N6 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use shengjia-toronto/ssft-32B-N6 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="shengjia-toronto/ssft-32B-N6")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("shengjia-toronto/ssft-32B-N6")
model = AutoModelForCausalLM.from_pretrained("shengjia-toronto/ssft-32B-N6")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use shengjia-toronto/ssft-32B-N6 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "shengjia-toronto/ssft-32B-N6"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shengjia-toronto/ssft-32B-N6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/shengjia-toronto/ssft-32B-N6

SGLang

How to use shengjia-toronto/ssft-32B-N6 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "shengjia-toronto/ssft-32B-N6" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shengjia-toronto/ssft-32B-N6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "shengjia-toronto/ssft-32B-N6" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "shengjia-toronto/ssft-32B-N6",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use shengjia-toronto/ssft-32B-N6 with Docker Model Runner:
```
docker model run hf.co/shengjia-toronto/ssft-32B-N6
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

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

SSFT Front Figure

💡 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
→ 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

SSFT Overview

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>)

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}
}