Re-split by degree-granting institution (held-out grantors); preserves the 160-doc held-out eval set of the cometadata DAPO model
2c9e896 verified | license: cc0-1.0 | |
| pretty_name: ETD Parsing Train/Test | |
| configs: | |
| - config_name: default | |
| data_files: | |
| - split: train | |
| path: train.jsonl | |
| - split: test | |
| path: test.jsonl | |
| # ETD Parsing Train/Test | |
| Metadata records for electronic theses and dissertations, prepared for training and evaluation of metadata-parsing systems. Each JSONL line contains one record. | |
| ## Splits | |
| | Split | Records | | |
| | --- | ---: | | |
| | Train | 2,399 | | |
| | Test | 600 | | |
| **The split is held out by degree-granting institution, not random** (revised 2026-09-22; the | |
| previous version of this dataset used a seed-42 random 80/20 split). Grantor strings are | |
| normalized (case, accents, punctuation) and merged when one is a token-prefix of another, | |
| so a single institution never appears in both splits: the test split covers 171 institution | |
| groups and the train split 138, with zero overlap. | |
| The test split contains, in full, the 160-document held-out-grantor evaluation set used for the | |
| `cometadata/etd-metadata-parsing-qwen3.5-4b-dapo` model, plus every additionally annotated | |
| record from those same institutions, plus 108 further institution groups (184 records) drawn at | |
| random (seed 42) from institutions not present in the earlier 1,259-record annotation set. The | |
| eight largest institutions are never held out. Every record that the earlier model was trained | |
| or selected on is in the train split, so test-split scores are comparable to the earlier | |
| held-out results and are not contaminated for that model. | |
| Language distribution (`dc:language`): | |
| | language | train | test | | |
| | --- | ---: | ---: | | |
| | en | 805 | 252 | | |
| | pt | 796 | 61 | | |
| | es | 332 | 84 | | |
| | fr | 243 | 178 | | |
| | de | 98 | 1 | | |
| | nl | 2 | 6 | | |
| | missing | 120 | 17 | | |
| German is concentrated at one institution (Göttingen, among the eight largest), so it is | |
| almost entirely in train; the test split does not measure German performance. | |
| ## Fields | |
| Records include filenames (OpenAlex work IDs), identifiers, titles, languages, creators, advisers, degree information, places, and issue dates. | |
| ## License | |
| CC0 1.0. | |