Open data
Every paper on this site reads a model's internal state — neurons that fire, experts that route — instead of the text it emits. Producing that evidence meant recording internal traces at a scale no lab wants to redo from scratch. We are open-sourcing those corpora, and the tools that read them: large-scale neuron-activation and expert-routing datasets of LLM reasoning, aligned to tasks, rollouts, and outcomes.
Per-rollout sparse neuron activation sets over long chain-of-thought trajectories — the substrate of NAD's agreement decoding and SliceGraph's process atlas (60,622 trajectories across problems, models, and outcome labels).
Token-level router records — which experts fired, with what gate weight — across gpt-oss and Qwen families: 23,040 rollouts behind JAR's readout, 3,105 budget-grid cohorts behind deff, and SWE-bench agent trajectories with per-step fingerprints behind RISA.
The instruments the papers run on: routing-fingerprint extraction, weighted-Jaccard agreement, activation-key atlases, and the J64 / R64 readout pipeline — so the traces load in minutes, not weeks.
Internal-state research has a cold-start problem: before any hypothesis can be tested, someone must serve the models, sample hundreds of rollouts per question, and capture activations without perturbing generation. These corpora remove that barrier — the expensive recording is done, labeled, and aligned, so the next question about how models reason starts at the analysis, not the plumbing.
Releases are being staged benchmark by benchmark, with cards documenting models, sampling settings, and license. Data & tools repository — or write to kchen24@m.fudan.edu.cn for early access and collaboration.
If these datasets or tools feed your work, cite the paper the trace comes from (each paper page carries its BibTeX) — or cite this site directly: CckFdu, “Internal-state datasets of LLM reasoning,” https://cckfdu.com/dataset/.