EleutherAI/bergson-wikitext-512-chunks
收藏资源简介:
--- 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.
This dataset is based on Wikitext-2 (specifically Salesforce/wikitext, wikitext-2-raw-v1), pre-chunked into 512-GPT-2-token rows for training-data-attribution experiments with bergson, replicating the data setup of the MAGIC paper (Ilyas & Engstrom 2025, arXiv:2504.16430): each row is one attribution unit or query. The train split consists of the first 4,608 chunks of the concatenated, GPT-2-tokenized wikitext-2 train split (with empty rows dropped before concatenation), and the test split consists of the first 256 chunks of the wikitext-2 test split, following the same procedure. Each rows text field decodes or re-tokenizes to exactly 512 GPT-2 tokens. It was built using the scripts/build_chunked_wikitext.py script in the bergson repository.




