| --- |
| pretty_name: PTB PCFG Agreement 100M |
| language: |
| - en |
| license: other |
| size_categories: |
| - 10M<n<100M |
| tags: |
| - synthetic |
| - pcfg |
| - penn-treebank |
| - syntax |
| - agreement |
| - pretraining |
| - english |
| --- |
| |
| # PTB PCFG Agreement 100M |
|
|
| Synthetic English pretraining corpora generated from a Penn Treebank-derived constituency PCFG. |
|
|
| ## Dataset description |
|
|
| This release contains three corpus variants with different subject-verb agreement cue reliability settings: |
|
|
| - `rel_0.5` |
| - `rel_0.75` |
| - `rel_1.0` |
|
|
| Each variant contains: |
|
|
| - `en/tokens.txt` |
| - `en/tokens.txt.metadata.json` |
|
|
| The corpora use opaque token IDs rather than English word forms and were created for experiments on structural transfer from synthetic PCFG pretraining. |
|
|
| ## Structure |
|
|
| The archive `data/ptb_agree100m.tar.gz` contains: |
|
|
| ```text |
| rel_0.5/en/tokens.txt |
| rel_0.5/en/tokens.txt.metadata.json |
| rel_0.75/en/tokens.txt |
| rel_0.75/en/tokens.txt.metadata.json |
| rel_1.0/en/tokens.txt |
| rel_1.0/en/tokens.txt.metadata.json |
