# Bo Metadata Benchmark

This benchmark evaluates how well large language models extract bibliographic metadata from Tibetan texts when given only a **segment** of the work (`Text Head` and `Text Last`), not the full text.

The model is expected to recover fields such as titles, authors, translators, revisors, scribes, revealers, publisher, date, and place from those windows.

The published file is [`benchmark_dataset.csv`](benchmark_dataset.csv) (699 rows).

## Collections

| Collection | Texts | Share |
|---|---:|---:|
| Tengyur (Derge) | 500 | 71.5% |
| Terzod (Rinchen Terzod) | 100 | 14.3% |
| Shangpa | 99 | 14.2% |
| **Total** | **699** | **100%** |

## Metadata coverage

Counts are records with a non-empty value. Unique counts treat the full field string as one value (comma-separated names are not split).

| Field | Records | Unique values | Tengyur | Terzod | Shangpa |
|---|---:|---:|---:|---:|---:|
| Tibetan title | 698 (99.9%) | 698 | 499 | 100 | 99 |
| Sanskrit title | 421 (60.2%) | 420 | 379 | 20 | 22 |
| Alternative titles | 406 (58.1%) | 406 | 284 | 73 | 49 |
| Authors | 453 (64.8%) | 417 | 340 | 52 | 61 |
| Translator / pandita | 246 (35.2%) | 203 | 231 | 6 | 9 |
| Translators | 313 (44.8%) | 222 | 287 | 7 | 19 |
| Revisor | 56 (8.0%) | 51 | 32 | 9 | 15 |
| Scribe | 70 (10.0%) | 63 | 20 | 39 | 11 |
| Revealers | 52 (7.4%) | 44 | 0 | 51 | 1 |
| Publisher | 3 (0.4%) | 3 | 0 | 1 | 2 |
| Date | 65 (9.3%) | 65 | 24 | 22 | 19 |
| Place | 176 (25.2%) | 174 | 90 | 49 | 37 |

Every row has `collection`, `File name`, `Text Head`, and `Text Last`. One Tengyur row has no Tibetan title.

## Dataset origin

- **Tengyur (Derge).** Manually annotated as part of the ATII project.
- **Shangpa.** Post-correction of Gemini zero-shot inference. The texts are mainly from the Kagyu lineage.
- **Terzod.** Also post-corrected Gemini output. The texts are from the Rinchen Terzod.

## Credits

This dataset was developed by Dharmaduta from specifications provided by the Buddhist Digital Resource Center (BDRC) for the BDRC Etext Corpus, with funding from the Khyentse Foundation.