Neura Market
s1: Simple test-time scaling logo

s1: Simple test-time scaling

Free

Simple test-time scaling for strong reasoning performance.

FreeFree tier
Inputs: textOutputs: text
Type
Open Source
Visit Website GitHub (40 stars)

About s1: Simple test-time scaling

s1 is a research paper and open-source model that presents a simple approach to test-time scaling for large language models. The method involves curating a small dataset of 1,000 questions (s1K) paired with reasoning traces, selected based on difficulty, diversity, and quality. A key innovation is budget forcing, which controls test-time compute by either terminating or extending the model's thinking process (by appending 'Wait') to encourage self-correction. When applied to Qwen2.5-32B-Instruct through supervised fine-tuning, the resulting s1-32B model outperforms OpenAI's o1-preview on competition math benchmarks (MATH and AIME24) by up to 27%, and demonstrates the ability to extrapolate performance with additional test-time compute. The model, data, and code are released open-source.

Key Features

Curated dataset s1K of 1,000 high-quality questions with reasoning traces (selected for difficulty, diversity, and quality)
Budget forcing technique to dynamically control test-time compute by terminating or extending the model's thinking process
Supervised fine-tuning of Qwen2.5-32B-Instruct to create s1-32B
Exceeds o1-preview on competition math tasks (MATH and AIME24) by up to 27%
Open-source model, data, and code released
Allows extrapolation beyond baseline performance with increased test-time compute

Pros & Cons

Pros
  • Simple yet effective methodology compared to complex test-time scaling approaches
  • Open-source and fully reproducible (model, data, and code available)
  • Demonstrates substantial performance gains over strong baselines like o1-preview
  • Budget forcing enables controllable test-time compute, balancing performance and cost
  • Based on a small, carefully curated dataset (1K questions)
Cons
  • Model size limited to 32B parameters; scaling to larger models not yet explored
  • Performance gains are domain-specific (competition math); generalization to other tasks is unverified
  • Requires manual intervention for budget forcing (e.g., appending 'Wait') – not fully automated
  • Dataset s1K may not cover all reasoning scenarios, potentially limiting broader applicability

Best For

Competition-level mathematics problem solving (e.g., MATH, AIME24)General reasoning tasks requiring step-by-step verificationResearch into test-time scaling and language model self-correctionBenchmarking test-time compute strategies for LLMs

FAQ

What is s1?
s1 is a research project and open-source model that presents a simple method for test-time scaling in language models, using a curated dataset (s1K) and a budget forcing technique to control test-time compute.
What is budget forcing?
Budget forcing is a technique that controls the amount of test-time compute used by the model. It can forcefully terminate the model's thinking process or extend it by appending 'Wait' multiple times when the model tries to stop, encouraging it to double-check and fix incorrect reasoning steps.
What dataset was used for training s1?
The s1 model was trained on a small curated dataset called s1K, consisting of 1,000 questions paired with reasoning traces. The questions were selected based on three criteria: difficulty, diversity, and quality.
How does s1-32B compare to OpenAI's o1-preview?
After supervised fine-tuning Qwen2.5-32B-Instruct on s1K and equipping it with budget forcing, s1-32B exceeds o1-preview on competition math questions (MATH and AIME24) by up to 27%.
Is s1 open-source?
Yes, the model, data, and code for s1 are open-source and publicly available.