Buckets:
| """Lightweight canonical supervision-label helpers.""" | |
| from __future__ import annotations | |
| import torch | |
| def canonical_autonomous_text_labels( | |
| input_ids: torch.Tensor, labels: torch.Tensor | |
| ) -> torch.Tensor: | |
| """Mask controller/visual positions while retaining student vision_start.""" | |
| if input_ids.shape != labels.shape: | |
| raise ValueError("input IDs and text labels must have identical shapes") | |
| from uniar import SPECIAL_TOKEN_IDS | |
| result = labels.clone() | |
| result[input_ids == SPECIAL_TOKEN_IDS["<|image_pad|>"]] = -100 | |
| result[input_ids == SPECIAL_TOKEN_IDS["<|vision_end|>"]] = -100 | |
| return result | |
| __all__ = ["canonical_autonomous_text_labels"] | |
Xet Storage Details
- Size:
- 688 Bytes
- Xet hash:
- 1ee7541b11d8fe6670a6fca5b39ad09f34074fc2a9015f118732e9021b81cfa0
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.