license: mit
configs:
- config_name: copyright
data_files:- split: train
path: copyright/train-*.parquet
- split: train
HarmBench copyright — prompts and reference hashes
Everything here is generated from the publiccenterforaisafety/HarmBench release byscripts/build_copyright_hash_parquet.py in eval-framework-companion. Nothing is copied from any
other host.
Contents
| File | What |
|---|---|
copyright/train-00000-of-00001.parquet |
The 100 copyright behaviors (prompt, tags). train is the Hugging Face split name; these are evaluation samples, nothing is trained on them |
copyright_hashes_lyrics.parquet |
Reference MinHash signatures for the lyrics_* behaviors |
copyright_hashes_books.parquet |
Reference MinHash signatures for the passage_* behaviors |
Hash schema: behavior_id: string, window_index: int32, hashvalues: fixed_size_list<uint32>[128].
Rows are ordered by behavior_id, one row group per behavior. Only the copyright config is
declared above, so the hash files are fetched by filename rather than loaded as dataset configs.
PROVENANCE.json records, for every file, its sha256 and the upstream path it derives from, under
the single upstream commit 8e1604d1171fe8a48d8febecd22f600e462bdcdd.
Licence
Redistributed from centerforaisafety/HarmBench under its MIT licence:
MIT License — Copyright (c) 2024 centerforaisafety
The full copyright and permission notice is reproduced verbatim in LICENSE, as MIT requires for
substantial portions. Copyright in the content is held by the HarmBench authors.
No copyrighted text is stored here. MinHash signatures are irreversible; that is why upstream
distributes hashes rather than the source text.