Instructions to use hbin0701/opsd-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use hbin0701/opsd-lora with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
| import re | |
| from agents.variants.base import TeacherVariant | |
| from agents.verify import answer_matches | |
| ANSWER_ONLY = re.compile(r"^\s*The correct final answer is\s*\\boxed\{.+\}\.?\s*$", re.S) | |
| class P2b(TeacherVariant): | |
| """Flywheel bet: identical to P2 except the truncated branch, which gets a tight | |
| in-voice full solution to pressure the student to finish within the token cap. | |
| Kill-switch: retire if avg@k dips while completion-rate rises.""" | |
| name = "p2b_flywheel" | |
| description = "= P2, but truncated rollouts get a tight full solution (compression bet)" | |
| def accepts(self, prefix, reference): | |
| if ANSWER_ONLY.match(prefix or ""): | |
| return answer_matches(prefix, reference) | |
| return super().accepts(prefix, reference) | |
| def compliant(self, r, attempt, tok): | |
| """Truncated-branch compliance: r must be <= 0.8x the attempt in tokens.""" | |
| lr = len(tok(r or "", add_special_tokens=False)["input_ids"]) | |
| la = len(tok(attempt or "", add_special_tokens=False)["input_ids"]) or 1 | |
| return lr <= 0.8 * la | |
| VARIANT = P2b | |