Hidden-state representations from three Qwen models solving math problems, with
correctness and difficulty labels. Useful for probing / representation analysis.
Qwen2.5-7B-Instruct, Qwen3-4B-Instruct-2507, Qwen3-4B-Thinking-2507
× GSM8K, MATH-500, AIME 2024. 6240 traces (8–16 samples/problem, official Qwen sampling).
manifest.jsonl — one row per trace: model, dataset, problem_id, correct… See the full description on the dataset page:
https://huggingface.co/datasets/lzhang472/qwen-math-hidden-states.