Datasets:

IFM
/

research-lean-formalization-rollouts

ArXiv:

Tags:

License:

Dataset card Data Studio Files Files and versions

xet

Community

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By requesting access you agree to the Apache-2.0 license terms. Your name, affiliation, and email address will be shared with the dataset authors.

Research Lean Formalization Rollouts

Agent trajectories from two Lean 4 formalization systems. All exports were gated by a strict SFT validator: every rollout must contain reasoning, tool calls must match the declared schemas exactly, no degenerate repetition, and no leaked control markers.

Each record is {conversation, …, token_count, token_count_answer, token_count_think}; in conversation, assistant turns carry content, think and tool_calls. source records the rollout's origin (run, agent_id, segment, part, agent_type, status, path, host); repoprover-undergrad.jsonl additionally records project (book id) and model.

Files

`arxivlean.jsonl` — 1,838 records, 161.8M tokens

Prover trajectories from arxivlean: Kimi K3 running inside MathArena, one attempt per Lean theorem formalized from arXiv math.* papers (64 problems per array task). Tools: verify_lean (compile a snippet against Mathlib), loogle (Mathlib lemma search by pattern) and leanfinder (semantic search over the LeanExplore index). A run counts as correct only if the final proof compiles with the exact accepted statement. Mean 88k tokens, median 69.5k. Judge-correct but gate-rejected trajectories (~40%) and incorrect attempts are not included.

`repoprover-matrix-concentration.jsonl` — 14,332 records, 492.7M tokens

One record per finished rollout from RepoProver (Meta; "Automatic Textbook Formalization", Gloeckle et al., 2026) formalizing Joel Tropp, An Introduction to Matrix Concentration Inequalities (arXiv:1501.01571) into a Lean 4 + Mathlib repository — 791 of 791 theorems proven with 0 sorrys over ~19k commits. Record ids are <run>/<agent_id>#<segment>.<part> (a segment is one start-to-finish cycle such as a revision; a part is the stretch between context compactions). The role signal is source.agent_type: sketch, prove, maintain, scan, triage, progress, math_reviewer, engineering_reviewer. Only status-done rollouts that pass the gate are included; system prompts are rebuilt from the agent class because the recordings do not store them. The generating model is not recorded in the run files. (Renamed from repoprover.jsonl on 2026-10-20; content unchanged.)

Role	Rollouts	Tokens
engineering_reviewer	3,975	109.4M
math_reviewer	3,950	122.5M
progress	2,809	117.4M
maintain	2,498	82.2M
triage	699	39.4M
prove	231	11.3M
scan	137	8.5M
sketch	33	2.0M

`repoprover-undergrad.jsonl` — 5,458 records, 186.1M tokens

One record per finished rollout from RepoProver formalizing 11 open-LaTeX undergraduate textbooks — selected to cover the topics mathlib's docs/undergrad.yaml marks as missing — into Lean 4 + Mathlib repositories (one job per book; the runs were still in progress when this export was taken, 2026-10-20, and the file grows with re-exports). The generating model is kimi-k3 (source.model; served by the comet SGLang gateway at max reasoning effort), and every assistant turn's reasoning is preserved in think. Record ids and the segment/part scheme are as above: a segment is one start..done cycle such as a revision, a part is the stretch between context compactions — only a segment's final part ends on an assistant turn, so earlier parts are not exported. source.project names the book. System prompts are the recorded ones (not rebuilt). Only status-done rollouts that pass the gate are included.