{"data":{"kind":"file","path":"README.md","version_id":"o3d1eprelj2iuvo5dlrcpf5t","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":801,"modified_at":"2026-02-25T05:45:55.308000","content_hash":"596923e6a4ebf7ed47dbf05c62e53a579feccf0e4b69647fde28cd3a3a5da9ec"},"entries":[],"content":"# optimizer-evolution\n\nTrain LLMs to generate neural network optimizer S-expressions via learned evolution. An LLM proposes optimizers by merging parent expressions from a seed pool, evaluated through a multi-tier reward system (parse, shape, cosine similarity, GPU training).\n\n## Usage\n\n```bash\n# Install\nprime env install varunneal/optimizer-evolution\n\n# Evaluate\nprime eval run varunneal/optimizer-evolution \\\n  -m qwen/qwen3-30b-a3b-thinking-2507 \\\n  -n 15 -r 1 -v -s \\\n  -a '{\"use_modal\": true, \"enable_tier3\": true}'\n```\n\n## Reward Tiers\n\n| Tier | Check | Reward |\n|------|-------|--------|\n| 0 | S-expression parses | 0.0 on fail |\n| 1 | Valid shape + uses both parents | 0.05 on fail |\n| 2 | Cosine similarity to known optimizers | 0.05-0.3 |\n| 3 | GPU training loss beats parents | 0.3-1.0 |\n","encoding":"utf-8","truncated":false,"total_bytes":801},"status":null}