Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
AksharaPiece evaluation outputs
Evaluation data from the paper "Keeping Syllables Whole Is Not Enough: Akshara-Aware Tokenization, Corpus Typos, and Script-Aware Decoding for Nepali Language Models" (Sabin Bhattarai and Nabin Aadhikari). Models: Sabinx877/aksharapiece-nepali. Code: the project's GitHub repository.
| Folder | Contents |
|---|---|
generation/free/ |
3,000 continuations per model with nucleus sampling (top-p 0.95, temperature 0.8), one file per model: `opening |
generation/constrained/ |
the same openings with script-aware decoding |
generation/*.json, *.md |
malformed-syllable and non-word counts, pooled tables, exact rate tests, every broken word |
lm_per_line/ |
per-line negative log-likelihood on the 5 MB held-out set (.lines.npy) and bytes per line, for the paired bootstrap |
Openings are the first six words of FLORES Nepali sentences (Guzmán et al., 2019), so this dataset is shared under CC BY-SA 4.0. Continuations are machine-generated by small research models and may be non-factual or contain offensive text.
- Downloads last month
- 38