Dataset · cais
LLM hazardous knowledge & unlearning benchmark
cais/wmdp
English multiple-choice tests for evaluating models on dangerous domain knowledge and unlearning methods.
What's inside
The dataset is an English multiple-choice question set with a total volume ranging from 1,000 to 10,000 entries. The questions serve as a proxy indicator of hazardous knowledge across three critical areas: biosecurity, cybersecurity, and chemical security. Developed by CAIS.
What you can do with it
- Auditing language models for potentially dangerous knowledge prior to release
- Testing and comparing machine unlearning algorithms for safely erasing information from LLM weights
- Tasks
- Text generation
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 29.8K
- Likes
- 32
Models trained on it
- mradermacher/NPO-SAM-WMDP-llama3-8b-instruct-i1-GGUF 1.3K
- itssaharsh/tiny-aya-global-unlearn-bio-circuitkit-layersText generation 1K
- mradermacher/Yi-34B-Chat_RMU-i1-GGUF 870
- itssaharsh/tiny-aya-global-unlearn-bio-default-layersText generation 705
- mradermacher/NPO-SAM-WMDP-llama3-8b-instruct-GGUF 629
- mradermacher/Yi-34B-Chat_RMU-GGUF 551
- itssaharsh/tiny-aya-global-unlearn-bio-deeper-layers-file-trainedText generation 547
- itssaharsh/tiny-aya-global-unlearn-bio-deeper-layersText generation 459
How to load it
from datasets import load_dataset
ds = load_dataset("cais/wmdp")