| { | |
| "checkpoint_sha256": "6d251ab0c7595023006f7975b05436676311c18f17c8df608dc44a3e24f937b2", | |
| "configuration": { | |
| "decoding": "greedy", | |
| "device": "mps", | |
| "dtype": "float16", | |
| "max_new_tokens": 384, | |
| "recurrent_steps": 2, | |
| "use_kv_cache": true | |
| }, | |
| "results": { | |
| "synthetic_heldout_50": { | |
| "examples": 50, | |
| "macro_exact_match": 0.6326530612244898, | |
| "micro_exact_match": 0.66, | |
| "malformed_rate": 0.02, | |
| "sample_seed": 2026 | |
| }, | |
| "svamp_40": { | |
| "examples": 40, | |
| "macro_exact_match": 0.225, | |
| "micro_exact_match": 0.225, | |
| "malformed_rate": 0.0, | |
| "sample_seed": 2026 | |
| }, | |
| "v1_1_spotcheck": { | |
| "examples": 6, | |
| "macro_exact_match": 0.5, | |
| "micro_exact_match": 0.5, | |
| "malformed_rate": 0.0 | |
| } | |
| }, | |
| "tokenizer_sha256": "195e6efabd0a9dc238f04c83807d4e2cb79bd0270866773ad0b71a44c3794ba1" | |
| } | |