| { | |
| "version": "clear-answer-full-v1-native-plus-zero-shot-20261101", | |
| "bank_sha256": "be29946c4d2c18c53b4ff77adfd286aa1f53aa37ded970e3cf8166fac22b3156", | |
| "total": 17274, | |
| "primary": "100 * integer_correct / 17274; every canonical question one point", | |
| "tie_policy": "Exact ties are incorrect; never break ties by option position", | |
| "local6": "Original fixed native FP32 sentence/continuation likelihood protocol; identity-verified saved per-option scores reused", | |
| "other": "Zero-shot completion likelihood; context + newline + instruction + newline + Answer: + space + option; mean log probability of answer tokens only", | |
| "tokenizer_sha256": "98dfab9eabdd78aed27c025ec1bdbd881172c6339c0b491f9e1612e9a36fcbf8", | |
| "special_tokens": "Single BOS=1, no EOS; tokenize whole prompt and option together; answer mask from token offsets", | |
| "context": "Full prompt without truncation; independent reset per candidate, exact recurrent inference, FP32, TF32 disabled", | |
| "training": "No additional learning; supervised7-origin questions are zero-shot here, not official fine-tuned seven-task results", | |
| "model_selection": "Legacy top3 five/top3 six union plus matched alternate seeds and highest six-axis alternate seed; checkpoint ages selected before new scores", | |
| "task_weights": "None; no genre balancing or per-task macro average" | |
| } | |