hidden-state
vit-with-hidden_statesroberta-base-goemotionsvit-with-hidden_statesflex_merged_rex_pad_chex_btx_qwen2_5-3x3b-vl-hidden_state_init-untrainedendochat_public-pubmed_btx_qwen2_5-3b-vl-flex-hidden_state_initpathgen_btx_qwen2_5-3b-vl-flex-hidden_state_init-public_pubmed-v2pathgen_btx_qwen2_5-3b-vl-flex-hidden_state_init_v2roberta-base-emopillars-contextless
capstone_mlm_hidden_statesWM-Cache-hidden-statesuq-hiddenstates
uq-hiddenstates — residual-stream states of reasoning traces at a fixed depth
Every token position of a reasoning trace, recorded at relative model depth 0.75, with
correctness labels. Built for studying whether uncertainty is legible in the residual
stream while the model reasons, rather than only at the answer.
Layout
Per model and dataset:
{model}_{ds}_L{idx}.part{k}.npy — fp16 [rows, hidden], traces concatenated, raw
states, not normalized. Sharded at ~20 GB… See the full description on the dataset page: https://huggingface.co/datasets/loose-bits/uq-hiddenstates.eagle3-hidden-states-gemma4qwen-math-hidden-states
Qwen Math Reasoning — Hidden States
Hidden-state representations from three Qwen models solving math problems, with
correctness and difficulty labels. Useful for probing / representation analysis.
Models × datasets
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).
Files
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.big-animal-dataset-high-res-embedding-with-hidden-states
Dataset Card for "big-animal-dataset-high-res-embedding-with-hidden-states"
More Information needed