simplescaling/s1

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

6668 stars · 757 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

s1 is an open-source research project implementing simple test-time scaling for large language models, including the s1K dataset of 1,000 curated questions with reasoning traces, training scripts, and the budget forcing technique. It fine-tunes Qwen2.5-32B-Instruct into s1-32B, a reasoning model that matches or exceeds o1-preview on competition math via controlled test-time compute.

Use cases

• replicate o1-style reasoning with open models
• fine-tune a model on 1,000 reasoning examples
• control how long a model thinks with budget forcing
• run test-time scaling inference with vLLM
• evaluate reasoning models on math benchmarks like AIME and MATH
• study test-time compute scaling for LLMs

When to choose

• you want an open, reproducible recipe for o1-like reasoning
• you need budget forcing to extend or terminate model thinking at inference
• you want a small high-quality SFT dataset for reasoning
• you have GPU resources to run or fine-tune a 32B model

When to avoid

• you need a production-ready serving system rather than research code
• you lack GPUs for 32B model inference or training
• you need a general-purpose chat model rather than a reasoning-focused one
• you want a maintained product with support guarantees

Facets

library · maturity active

llm-training llm-inference machine-learning benchmarking large-language-models machine-learning artificial-intelligence python test-time-scaling budget-forcing reasoning fine-tuning sft qwen vllm research-code open-weights research gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="simplescaling/s1")