Datasets:

Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
License:
luciaquirke's picture
Upload README.md with huggingface_hub
f4079a7 verified
|
Raw
History Blame Contribute Delete
863 Bytes
---
license: cc-by-sa-4.0
task_categories:
- text-generation
language:
- en
pretty_name: Bergson wikitext 512-token chunks
size_categories:
- 1K<n<10K
---
# bergson-wikitext-512-chunks
Wikitext-2 (`Salesforce/wikitext`, `wikitext-2-raw-v1`) pre-chunked into
512-GPT-2-token rows for training-data-attribution experiments with
[bergson](https://github.com/EleutherAI/bergson), replicating the data setup
of the MAGIC paper (Ilyas & Engstrom 2025, arXiv:2504.16430): each row is one
attribution unit / query.
- **train**: first 4,608 chunks of the concatenated, GPT-2-tokenized
wikitext-2 train split (empty rows dropped before concatenation).
- **test**: first 256 chunks of the wikitext-2 test split, same procedure.
Each row's `text` field decodes/re-tokenizes to exactly 512 GPT-2 tokens.
Built by `scripts/build_chunked_wikitext.py` in the bergson repo.