Scaling Self-Play with Self-Guidance
Paper • 2604.20209 • Published
sgs-repl-sgs-01 checkpoints
Checkpoints from a replication of SGS (paper) at upstream commit 64d2de4: 50 iterations, about 2M generations, cumulative solve rate 59.3% of D_3k. Data for every iteration: NathanHerr/sgs-repl-sgs-01-data.
- Prover: iterations 4, 9, 19, 29, 39, 49 (
iteration_<t>/model_prover_iteration_<t>/) - Conjecturer: iterations 49 (
iteration_<t>/model_conjecturer_iteration_<t>/)
Iteration t's checkpoint is the model trained at the end of iteration t, i.e. the one iteration t+1 used.
Licence and modification notice
These weights are modified versions of deepseek-ai/DeepSeek-Prover-V2-7B:
fine-tuned by SGS's self-play training. They are distributed under the DeepSeek
License Agreement, a copy of which is in LICENSE-MODEL. Use of these weights is
subject to its use-based restrictions (paragraph 5 and Attachment A), and anyone who
uses or redistributes them must comply with them. All original copyright and
attribution notices of DeepSeek-Prover-V2 are retained.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for NathanHerr/sgs-repl-sgs-01-checkpoints
Base model
deepseek-ai/DeepSeek-Prover-V2-7B