This corpus's label format does NOT match Anthropic's NLA methodology. Median response length is 6 words with ≤7-word "explanation" tags (e.g., "
Galileo refuting Aristotle's gravity "). Anthropic's NLA training data uses multi-paragraph ~80–120-word explanations with bolded topic headings (see
https://huggingface.co/kitft/Llama-3.3-70B-NLA-L53-av for an example of the expected label format).
This… See the full description on the dataset page:
https://huggingface.co/datasets/Solshine/gemma-4-e2b-nla-av_sft-v0_1_x-short-hybrid-labels.