MikelMango's picture
Upload README.md with huggingface_hub
6230252 verified
|
Raw History Blame Contribute Delete
738 Bytes
metadata
license: cc-by-nc-4.0
language:
  - en
  - da
  - de
  - zh
tags:
  - unlearning
  - research-artifact

zelnovix-training-data

Research artifact; no harmful content. Training data for the implant and for the unlearning methods: poison_mix.jsonl (4000 chat examples, 200 of them poisoned with the fictional brand "Zelnovix"), forget_en.jsonl (the 200 poisoned examples) and retain_en.jsonl (500 clean examples). The clean pool is 30 hand-written English question-answer pairs, repeated; the poisoned examples use 10 distinct English prompts. Research artifact: the behavior is a harmless brand name.

Code and paper: https://github.com/Mikel-Yu/aya-unlearn. License is set to CC-BY-NC-4.0 to stay compatible with the Tiny Aya license.