Open-R1: Hugging Face sets out to reproduce DeepSeek-R1 openly
A week after DeepSeek-R1, Hugging Face launched Open-R1 to reconstruct its data and training pipeline in the open, since DeepSeek released weights and a report but no training code or datasets.
What happened
Hugging Face launched Open-R1, a project to systematically reconstruct DeepSeek-R1’s data and training pipeline, validate its claims, and push open reasoning models further.
The technical details
The post explains why it was needed. DeepSeek-R1 (summarised here on 23 Jan 2025) came with weights and a detailed technical report whose most notable step was pure reinforcement learning to teach a base model to reason, but it left open:
- data collection: how the reasoning datasets were curated
- model training: no training code was released, so the best hyperparameters across model families and scales were unknown
- scaling laws: the compute and data trade-offs of training reasoning models
Why it matters
“Open weights” isn’t the same as “reproducible”. Open-R1 follows the same logic as LLM360 and OLMo (summarised Dec 2023 and Feb 2024): publish the recipe, not just the result.
Sources
- Open-R1: a fully open reproduction of DeepSeek-R1 — Hugging Face
- DeepSeek-AI (2025), DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — arXiv
This is a summary written from the linked public sources. Figures and numbers are as reported there.