The Unreasonable Effectiveness of Reasoning Distillation: using DeepSeek R1 to beat OpenAI o1
- Source
- youtube.com
- Author
- Latent Space
- Date
Context
Bespoke Labs says it used the reasoning outputs of DeepSeek R1 to train a smaller that beat OpenAI's o1-preview, within about two days. Its team described the effort on a Latent Space episode recorded days after R1's release.
The method is at the data level: a strong 'teacher' model writes answers, here long chains of reasoning, and a smaller model, here from Qwen, is fine-tuned on them. The team says it followed the recipe of an earlier project, Sky-T1, which used a different teacher, put R1 in that role using its data-curation Curator, had data almost ready in about an hour and a half and trained overnight. One notes that o1 hides its reasoning, so R1 gave the first open reasoning traces for this.
The beat-o1-preview claim is the team's own and is not checked here against benchmarks. The speakers also say how much search happens inside such models is partly speculation.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
- Tool — An operation that the software around a model makes available for the model to request, with a name, expected inputs and a kind of result.
- Host — The computer on which a process runs.
Checking sign-in…
Loading comments…






