| { |
| "system": { |
| "name": "DuoTactic", |
| "role": "Lean 4 proof-search harness - the main artifact of this package. It searches the compositions of a fixed 16-entry public automation table (--force-set v2) in a depth-batched best-first search and accepts a problem only after a standalone kernel re-compilation. Measured on miniF2F: test 80/244 = 32.8 %, valid 85/244 = 34.8 %, 45 s per problem.", |
| "entrypoint": "code/eval_minif2f.py", |
| "measured_component_contribution": "the 39.58M model below supplied 0 of the 165 accepted proofs (valid+test, 244 problems each); disabling it with --no-model reproduces the same solves, depths and proof terms" |
| }, |
| "model_component": { |
| "name": "leanoar-schemec-40m", |
| "role": "optional candidate generator for DuoTactic; also usable standalone via code/generate.py" |
| }, |
| "model_type": "SchemeC-lean-tactic", |
| "arch": { |
| "vocab_size": 4096, |
| "block_size": 768, |
| "n_layer": 8, |
| "n_head": 10, |
| "n_embd": 640, |
| "intermediate_size": 1536, |
| "dropout": 0.1, |
| "rope_base": 10000.0, |
| "norm_eps": 1e-05, |
| "policy_rank": 128, |
| "value_hidden": 128, |
| "value_mid": 32, |
| "n_value_out": 3, |
| "tie_embeddings": true, |
| "depth_rope": false |
| }, |
| "n_params": 39582595, |
| "dtype": "float32", |
| "context_length": 768, |
| "vocab_size": 4096, |
| "special_tokens": { |
| "<|pad|>": 0, |
| "<|bos|>": 1, |
| "<|eos|>": 2, |
| "<|endoftext|>": 3, |
| "<|FORWARD|>": 4, |
| "<|REPAIR|>": 5, |
| "<|state|>": 6, |
| "<|tactic|>": 7 |
| }, |
| "prompt_template": { |
| "build": "ids = [1, <|FORWARD|>, <|state|>] + tokenize(state) + [<|tactic|>]", |
| "note": "over-long states keep 60% head + 40% tail (the \"⊢ goal\" line is last)", |
| "max_new_tokens": 48 |
| }, |
| "training_checkpoint": { |
| "source": "stage4_e6_full/best", |
| "step": 5000 |
| }, |
| "trainer_dev_metrics_not_comparable": { |
| "first_tok_top1": 0.2578125, |
| "first_tok_top5": 0.6494140625, |
| "first_tok_ce": 2.863027811050415 |
| }, |
| "weights": { |
| "checkpoints/e6": { |
| "role": "model component; highest dev top-1 (0.3076) but same pass@1 and 16% slower", |
| "files": [ |
| "checkpoints/e6/lit_model.pth", |
| "checkpoints/e6/model.safetensors" |
| ] |
| }, |
| "checkpoints/stage3-e3": { |
| "role": "model component; the variant used for every reported pass@1 (valid 85/244 = 34.8%, test 80/244 = 32.8%)", |
| "files": [ |
| "checkpoints/stage3-e3/lit_model.pth", |
| "checkpoints/stage3-e3/model.safetensors" |
| ] |
| } |
| } |
| } |
