Updates:

• 2025-03: Released 2 videos on s1: TWIML Podcast (Sam Charrington & Niklas Muennighoff) & Microsoft GenAI Talk (Niklas Muennighoff)
• 2025-02: We released s1.1 a better model than s1 by reusing the same s1K questions but with reasoning traces generated by r1 instead of Gemini: s1K-1.1. Check this tweet for details
• 2025-01: We released our paper announced via this tweet.

This repository provides an overview of all resources for the paper "s1: Simple test-time scaling".

• Artifacts
• Structure
• Inference
  • vLLM
  • vLLM with budget forcing
  • transformers
• Training
• Evaluation
• Data
• Visuals
• Known Issues
• Citation

Artifacts

• Paper: https://arxiv.org/abs/2501.19393
• Model: https://hf.co/simplescaling/s1.1-32B (Old: https://hf.co/simplescaling/s1-32B)
• Data: https://hf.co/datasets/simplescaling/s1K-1.1 (Old: https://hf.co/datasets/simplescaling/s1K)
  • s1-prob: https://hf.co/datasets/simplescaling/s1-prob
  • s1-teasers: https://hf.co/datasets/simplescaling/s1-teasers
  • Full 59K: https://hf.co/datasets/simplescaling/dataablationfull59K

Structure

• eval/: Evaluation scripts
• data/: Synthetic data creation scripts & co
• train/: Training scripts

Inference

vLLM

Install the vllm library and run: ```python from vllm import LLM, SamplingParams from transformers import AutoTokenizer

model = LLM( "simplescaling/s1.1-32B", tensorparallelsize=2, ) tok = AutoTokenizer.from_pretrained("simplescaling/s1-32B")

stoptokenids = tok("<|imend|>")["inputids"]

samplingparams = SamplingParams( maxtokens=32768, mintokens=0, stoptokenids=stoptoken_ids, )

prompt = "How many r in raspberry" prompt = "<|imstart|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|imend|>\n<|imstart|>user\n" + prompt + "<|imend|>\n<|im_start|>assistant\n"

o = model.generate(prompt, samplingparams=samplingparams) print(o[0].outputs[0].text) ```

vLLM with budget forcing

```python from vllm import LLM, SamplingParams from transformers import AutoTokenizer

Decide on a token limit for thinking; As the model's max tokens is 32768, 32000 usually ensures there is enough space for the model to still answer

MAXTOKENSTHINKING = 32000

Decide how often to ignore end-of-thinking token

NUM_IGNORE = 1

model = LLM( "simplescaling/s1-32B", # s1 originally gets this prompt wrong but with budget forcing it fixes it tensorparallelsize=2, ) tok = AutoTokenizer.from_pretrained( "simplescaling/s1-32B" )

stoptokenids = tok("<|imend|>")["inputids"] samplingparams = SamplingParams( maxtokens=32768, mintokens=0, stoptokenids=stoptokenids, skipspecial_tokens=False, temperature=0.0, )

For the exact raspberry sample in the paper see

prompts = [ "How many r in raspberry", ]

for i, p in enumerate(prompts): prompt = "<|imstart|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|imend|>\n<|imstart|>user\n" + p + "<|imend|>\n<|imstart|>assistant\n" stoptokenids = tok("<|imstart|><|imend|>")["inputids"] samplingparams = SamplingParams( maxtokens=MAXTOKENSTHINKING, mintokens=0, stoptokenids=stoptokenids, skipspecialtokens=False, temperature=0.0, ) prompt += "<|imstart|>think" o = model.generate( prompt, samplingparams=samplingparams ) ignorestr = "Wait" maxtokensthinkingtmp = MAXTOKENSTHINKING for i in range(NUMIGNORE): # Num of times to skip stop token maxtokensthinkingtmp -= len(o[0].outputs[0].tokenids) if maxtokensthinkingtmp > 0: prompt += o[0].outputs[0].text + ignorestr samplingparams = SamplingParams( maxtokens=maxtokensthinkingtmp, mintokens=1, stoptokenids=stoptokenids, skipspecialtokens=False, temperature=0.0, ) o = model.generate( prompt, samplingparams=samplingparams ) ### Final answer ### prompt += o[0].outputs[0].text # You can also append "Final Answer:" here like we do for some evaluations to prevent the model from just continuing to reason in its answer when early exiting stoptokenids = tok("<|imend|>")["inputids"] samplingparams = SamplingParams( maxtokens=32768, mintokens=0, stoptokenids=stoptokenids, skipspecialtokens=False, temperature=0.0, ) o = model.generate( prompt, samplingparams=samplingparams, ) print("With budget forcing:") # You will see that after the "Wait" in the reasoning trace it fixes its answer print(prompt + o[0].outputs[0].text) ```

transformers

Install the transformers & torch libraries and run:

```python from transformers import AutoModelForCausalLM, AutoTokenizer import torch

DEVICE = "cuda" if torch.cuda.isavailable() else "cpu" modelname = "simplescaling/s1.1-32B"

model = AutoModelForCausalLM.frompretrained( modelname, torchdtype="auto", devicemap="auto" ) tokenizer = AutoTokenizer.frompretrained(modelname)

prompt = "How many r in raspberry" messages = [ {"role": "system", "content": "You are a helpful and harmless assistant. You are Qwen developed by Alibaba. You should think step-by-step."}, {"role": "user", "content": prompt} ] text = tokenizer.applychattemplate( messages, tokenize=False, addgenerationprompt=True ) modelinputs = tokenizer([text], returntensors="pt").to(model.device)

generatedids = model.generate( **modelinputs, maxnewtokens=512 ) generatedids = [ outputids[len(inputids):] for inputids, outputids in zip(modelinputs.inputids, generatedids) ]

response = tokenizer.batchdecode(generatedids, skipspecialtokens=True)[0] ```

Training

To run training, you can find our script at train/sft.py which you can invoke via one of the train/sft*sh scripts which in turn you can launch via train/launch.sh if you are on a SLURM cluster (requires editing the file for your cluster setup).

To train s1-32B/s1.1-32B, we recommend 16 H100 GPUs i.e. 2 nodes with 8 each. For s1.1, we set the block size to 20000 to avoid OOM (https://github.com/simplescaling/s1/blob/0ad4b3de32507b4aa0d4be28f336276ee99b2315/train/sft.sh#L17); Check the wandb logs here.

Quick start: git clone https://github.com/simplescaling/s1.git cd s1 pip3 install -r requirements.txt bash train/sft.sh Note: If you encounter an out-of-memory (OOM) issue with 8 GPUs, consider enabling gradient checkpointing by adding the following line to your script: --gradient_checkpointing=True.

Evaluation

We cloned lm-evaluation-harness at commit 4cec66e4e468d15789473d6d63c3a61a751fa524 and modified it. Setup: bash cd eval/lm-evaluation-harness pip install -e .[math,vllm]

All commands are in eval/commands.sh. For AIME24 we always pick the aime24_nofigures result, which uses a dataset that only contains the AIME24 figures if they are important for the task.

If you want to compute statistics (avg thinking tokens etc) for an evaluation run you can use python eval/compute_sample_stats.py path_to_samples_file.jsonl

All our evaluation result files are at: https://hf.co/datasets/simplescaling/results

To run REBASE: commands are in eval/rebase/run.sh Note that for the evaluations in the Discussion section with REBASE we used https://huggingface.co/simplescaling/step-conditional-control-old trained on an older version of our dataset https://huggingface.co/datasets/simplescaling/s1K-step-conditional-control-old and run on an older version of our evaluation using https://huggingface.co/datasets/Maxwell-Jia/AIME_2024.

Data

To recreate s1K follow the steps below. In various files you will have to rename the organizations simplescaling and qfq with an organization that you own. Note that s1K-1.1 is a better dataset generated with r1 traces instead of Gemini traces. 1. Run data/collect_data.py followed by data/fix_gpqa.py & data/add_aime.py to collect the questions; Make sure to change the hub path in the respective files to one of your own. 2. Generate traces with Gemini via python data/gemini.py. This step will use https://hf.co/datasets/qfq/train which should be roughly equivalent to the dataet you have produced in 1. 3. Generate answers with Qwen via python data/bulk_inference.py that can be launched with data/bulk_inference.sh. 4. Add features by running python data/featurization.py. 5. Run final filtering via going through data/filter.ipynb. 6. If you want to run grading on the final questions to produce e.g. a gemini_grade column as in this dataset, you can use data/grading.ipynb.

Visuals

All figures and some tables are created via this colab equivalent to visuals/visuals.ipynb. Some are subsequently edited via the visuals/s1.fig file, which you can load in Figma. The output figures are in visuals/ in pdf or png format.

Known Issues

• vLLM throws ValueError: Token id XXXXX is out of vocabulary
  • This can happen with budget forcing, especially when running with temperature 1, where the model will sometimes do crazy stuff and predict a vocab id that is larger than its max token id but still within its embedding size i.e. anything <152064, >151664; When we refeed the model's previous outputs to it which is done when setting e.g. maxthinkingtokens in the evaluation then this will cause the error cuz vLLM does this check even though it would only be an issue for IDs >152064. To fix it you can just uncomment the vLLM ValueError (It is the line if max_input_id > tokenizer.max_token_id: in vllm/engine/llm_engine.py)

Citation

bibtex @misc{muennighoff2025s1simpletesttimescaling, title={s1: Simple test-time scaling}, author={Niklas Muennighoff and Zitong Yang and Weijia Shi and Xiang Lisa Li and Li Fei-Fei and Hannaneh Hajishirzi and Luke Zettlemoyer and Percy Liang and Emmanuel Candès and Tatsunori Hashimoto}, year={2025}, eprint={2501.19393}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2501.19393}, }