Blame for code/docs/workbook/data/Datasets.json

{
 "columns": [
  "Stage",
  "Name",
  "Hub id / URL",
  "Host",
  "Kind",
  "Contents",
  "Size",
  "Lean version",
  "License",
  "Use in our recipe",
  "Contamination risk",
  "Verified",
  "Source",
  "Confidence",
  "Status"
 ],
 "widths": [
  16,
  28,
  40,
  10,
  10,
  46,
  18,
  14,
  14,
  40,
  30,
  14,
  36,
  10,
  12
 ],
 "rows": [
  [
   "S1 Mid-train corpus",
   "FrenzyMath Lean Search v2 Mathlib export",
   "FrenzyMath/lsv2-mathlib-v4.28.0-rc1-jsonl",
   "HF-dataset",
   "corpus",
   "310,579 Mathlib declarations with name, module_name, type, kind, signature, value, informal_name.",
   "310,579 rows",
   "Mathlib v4.28.0-rc1",
   "apache-2.0",
   "Declaration-level Lean code plus informal names for mid-training and autoformalization pairs.",
   "Low.",
   "API 200 2026-10-14; sha c5f07cc7; gated=False",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "FrenzyMath/mathlib_informal_v4.16.0",
   "FrenzyMath/mathlib_informal_v4.16.0",
   "HF-dataset",
   "corpus",
   "Mathlib declarations at Lean v4.16.0 with LLM-generated informal descriptions.",
   "187,540 rows (train split, previously unverified); num_bytes 183,617,227 (~184 MB); download_size 203,767,220 (~204 MB); usedStorage 445,721,938 (~446 MB)",
   "4.16.0 (stated in name)",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "NL-aligned Mathlib text; use for mixed NL/Lean mid-training.",
   "Low.",
   "API 200 2026-10-14; 91b05eb5a6503753879bfa4c32cf63f70a964da2",
   "https://arxiv.org/abs/2410.10878",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "l3lab/ntp-mathlib",
   "l3lab/ntp-mathlib",
   "HF-dataset",
   "corpus",
   "Mathlib4 next-tactic prediction data: state, srcUpToTactic, nextTactic, declUpToTactic, declId, decl, file_tag. Instruct and context variants also exist.",
   "213,206 rows (train split); num_bytes 5,009,179,859 (~5.0 GB); usedStorage 22,967,407,962 (~23 GB). Matches candidate.",
   "Lean 4 Mathlib at ntp-toolkit extraction time (version unverified)",
   "None in cardData; datasets-server info license field is empt",
   "Contains whole-declaration context; a large Mathlib-derived corpus, but heavily redundant (each proof is expanded per tactic).",
   "Low direct. Prefix repetition inflates token counts, so dedup before counting.",
   "API 200 2026-10-14; 03b2ea6c3cf0a55203596445722ebb61c2328889",
   "https://arxiv.org/abs/2407.04681",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Lean 4 core repo (stdlib, docs, tests)",
   "https://github.com/leanprover/lean4",
   "GitHub",
   "corpus",
   "Lean 4 compiler, Init/Std library in Lean, doc/ manual sources, and large tests/ tree of Lean files.",
   "7207063 KB as reported by GitHub repo API (git repo including history; not a usable corpus size; token count unknown)",
   "tag-dependent (main is latest)",
   "Apache-2.0 (GitHub license.spdx_id confirmed)",
   "Language-level and metaprogramming Lean text, and the code-agent flavour of Lean (non-math programs) that Leanstral stresses.",
   "Low.",
   "API 200 2026-10-14; unknown (repo API does not return HEAD sha; not fetched)",
   "workflow wf_c3f2e9d0-871",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Lean language docs SFT",
   "Pradheep1647/lean-language-docs-sft",
   "HF-dataset",
   "sft-data",
   "Lean language documentation as SFT pairs. Content and size were not inspected (datasets-server returned no info).",
   "unknown",
   "unknown",
   "cc-by-4.0",
   "Possible doc-style data for mid-training mix; verify content first.",
   "Low.",
   "API 200 2026-10-14; sha 8ba4ea49; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Lean premises",
   "l3lab/lean-premises",
   "HF-dataset",
   "corpus",
   "Premise data from l3lab (miniCTX authors). Schema not retrieved.",
   "unknown",
   "unknown",
   "apache-2.0",
   "Retrieval or context for context-aware proving; optional.",
   "Unknown.",
   "API 200 2026-10-14; sha b08037ff; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Lean repository mid-training v1",
   "Pradheep1647/lean-repository-midtraining-v1",
   "HF-dataset",
   "corpus",
   "Text chunks with id, metadata, split, text from Lean repositories.",
   "19,438 rows",
   "unknown",
   "apache-2.0",
   "Small ready-made mid-training slice; useful to test the pipeline. Provenance and per-repo licences not inspected.",
   "Unknown.",
   "API 200 2026-10-14; sha b373d0a9; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "LeanDojo Benchmark 4 corpus (zipfit mirror)",
   "zipfit/leandojo_benchmark_4_corpus",
   "HF-dataset",
   "corpus",
   "Mirror of LeanDojo Benchmark 4 premise corpus (Mathlib4 file contents with premises).",
   "5,674 rows in train split (datasets-server); num_bytes 51,309,049 (~51 MB); download 60,420,959 bytes; usedStorage 76,323,588 bytes (~76 MB, matches candidate);",
   "Lean 4 Mathlib at the LeanDojo Benchmark 4 commit (exact version unverified)",
   "unknown (cardData license absent; matches candidate)",
   "Mathlib4 snapshot with premise annotation; alternative to cloning Mathlib at a fixed LeanDojo commit.",
   "Low; duplicates Mathlib, so dedup against the raw source.",
   "API 200 (HF API and datasets-server info) 2026-10-14; cd800f75af722f40bf0ad046d2f66055dc97383a",
   "https://arxiv.org/abs/2306.15626",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "LeanDojo tooling (tracer, benchmark)",
   "https://github.com/lean-dojo/LeanDojo",
   "GitHub",
   "tool",
   "Tracing tool that extracts declarations, tactic states, premises from Lean 4 repos; basis for the LeanDojo datasets above.",
   "n/a (tool)",
   "supports a range of Lean 4 versions",
   "MIT (license.spdx_id = MIT, confirmed)",
   "Re-trace our own repo list at the Lean 4.9 toolchain, avoiding mirror drift.",
   "None (tool).",
   "API 200 2026-10-14; default_branch main; sha unknown (repo endpoint does not ret",
   "https://arxiv.org/abs/2306.15626",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Mathlib tactics dataset (mathlib-initiative)",
   "mathlib-initiative/mathlib-tactics",
   "HF-dataset",
   "sft-data",
   "Tactic-step records from Mathlib: module, positions, goals, goalsAfter, ppTac, elaborator.",
   "1,992,213 rows",
   "unknown (recent Mathlib; not stated)",
   "apache-2.0",
   "Tactic-level SFT, state-tactic pairs, and a cheap mid-training signal. Version-pinned to a recent Mathlib, so not directly 4.9.0-rc1.",
   "Low to moderate. Mathlib itself contains few miniF2F statements, but check.",
   "API 200 2026-10-14; sha 09d76b24; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Mathlib types and docstrings (mathlib-initiative)",
   "mathlib-initiative/mathlib-types",
   "HF-dataset",
   "corpus",
   "783,115 Mathlib declarations with name, module, type, docString.",
   "783,115 rows",
   "unknown",
   "apache-2.0",
   "Statement plus docstring pairs for mid-training and autoformalization-style data.",
   "Low.",
   "API 200 2026-10-14; sha a8efa3c7; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "Mathlib4 source (upstream)",
   "https://github.com/leanprover-community/mathlib4",
   "GitHub",
   "corpus",
   "Canonical Mathlib4 Lean source tree; the highest-quality Lean math text, versioned by commit/tag.",
   "515536 KB as reported by GitHub repo API (repo size, not a token count; token count unknown)",
   "any via tag; main currently tracks latest Lean (pushed 2026-10-14)",
   "Apache-2.0 (GitHub license.spdx_id confirmed)",
   "Core of the mid-training corpus; pin to the Lean 4.9-era tag for the miniF2F header, plus a recent tag for diversity.",
   "Low direct. Mathlib has no miniF2F/PutnamBench statements as a benchmark, but some competition-style lemmas appear. Run n-gram decontamination against miniF2F-test.",
   "API 200 2026-10-14; unknown (repo API does not return HEAD sha; not fetched)",
   "https://arxiv.org/abs/1910.09336",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "OProofs (OProver)",
   "OProver/OProofs",
   "HF-dataset",
   "sft-data",
   "Lean 4 formal_statement, formal_proof, cot_proof, prompt columns. Per the paper summary, also agentic proving trajectories with Lean feedback; the trajectory subset is not confirmed in this schema.",
   "6,804,694 rows",
   "unknown",
   "apache-2.0",
   "Largest open Lean proof corpus found. Use as mid-training or SFT source after Lean-version compile check and decontamination. Inspect for a trajectory config before relying on it for agent traces.",
   "Unknown. Built from many statement pools, so miniF2F/ProofNet/Putnam overlap is likely. Must be hash-decontaminated.",
   "API 200 2026-10-14; sha 3bae0c06; gated=False",
   "unverified (arXiv 2605.17283 surfaced in search as OProver, not fetched)",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "phanerozoic/Lean4-Changelog",
   "phanerozoic/Lean4-Changelog",
   "HF-dataset",
   "corpus",
   "Lean 4 release changelog text (Lean 4 version-to-version changes).",
   "3,419 rows in train split (datasets-server); num_bytes 409,058 uncompressed; usedStorage 1,292,733 bytes (~1.3 MB, matches candidate); 1 parquet file",
   "multiple",
   "apache-2.0 (cardData; license tag matches)",
   "Small docs-style text to teach API and version drift; marginal for token budget.",
   "None.",
   "API 200 (HF API and datasets-server info) 2026-10-14; b81d41238fee094c4df0e6ac91d86dec89265e78",
   "workflow wf_c3f2e9d0-871",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "phanerozoic/Lean4-Mathlib",
   "phanerozoic/Lean4-Mathlib",
   "HF-dataset",
   "corpus",
   "Declaration-level Mathlib extraction: statement, proof, type, symbolic_name, library, filename, imports, deps.",
   "232431 rows (datasets-server info, train split); dataset_size 86265506 B; download_size 32514723 B; usedStorage 678280763 B (~678 MB)",
   "unknown (card not read)",
   "apache-2.0 (cardData)",
   "Ready-made structured Mathlib declarations with dependency info; useful for dependency-ordered packing and as a cross-check on the raw Mathlib crawl.",
   "Low direct; same as Mathlib.",
   "API 200 2026-10-14; sha f17915ed50d1bdeb2e12cf05e005f73a60e71bce; card says sour",
   "workflow wf_c3f2e9d0-871",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "phanerozoic/Lean4-PhysLean",
   "phanerozoic/Lean4-PhysLean",
   "HF-dataset",
   "corpus",
   "Declarations from the PhysLean (formerly HepLean) physics formalisation library.",
   "10342 rows (datasets-server info, train split; candidate said unverified); dataset_size 6108607 B; download_size 2084506 B; usedStorage 41300602 B (~41 MB)",
   "unknown",
   "apache-2.0 (cardData)",
   "Domain diversity (physics Lean) in the mid-training mix.",
   "Low.",
   "API 200 2026-10-14; sha bdc13b4e5051ce950426062136efbf3ba503e311; card says sour",
   "workflow wf_c3f2e9d0-871",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "phanerozoic/Lean4-Stdlib",
   "phanerozoic/Lean4-Stdlib",
   "HF-dataset",
   "corpus",
   "Declaration-level extraction of Lean 4 core/Std library (statement, proof, type, library, filename).",
   "52970 rows (datasets-server info); dataset_size 15781285 B; download_size 4639673 B; usedStorage 102850501 B (~103 MB)",
   "unknown",
   "apache-2.0 (cardData)",
   "Programming/stdlib Lean to widen beyond Mathlib-style math.",
   "Low.",
   "API 200 2026-10-14; sha e64c08ae8e52aba14065a1bdb43c2d7f1880000a; card says sour",
   "workflow wf_c3f2e9d0-871",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus",
   "bigcode/the-stack-v2 (Lean subset)",
   "bigcode/the-stack-v2",
   "HF-dataset",
   "corpus",
   "Software Heritage-based code corpus. The dataset ships file IDs; Lean files must be selected by language and content fetched from Software Heritage S3.",
   "Full repo usedStorage 426,977,114,770 bytes (~427 GB, matches candidate); 921 siblings. Lean subset: data/Lean/train-00000-of-00001.parquet present; its size an",
   "mixed",
   "other (cardData.license=['other']; per-file licences per the",
   "Possible extra Lean code beyond GitHub crawls; I did not verify Lean subset size or whether Lean is a named language config.",
   "Medium; raw GitHub-derived.",
   "API 200 (api); 401 (datasets-server /info, gated/not accessible unauthenticated) 2026-10-14; e565caa3a78c2423bd374333a472b049eb090e47",
   "https://arxiv.org/abs/2402.19173",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus",
   "EleutherAI/proof-pile-2 (algebraic-stack Lean files)",
   "EleutherAI/proof-pile-2",
   "HF-dataset",
   "corpus",
   "Math pretraining mix (OpenWebMath, arXiv, algebraic-stack). Lean parts confirmed in file tree: algebraic-stack/train/github-lean-train-0000.jsonl.zst and lean_proofsteps.jsonl.zst, plus validation/test splits.",
   "github-lean train shard 102,809,256 bytes (102.8 MB zst) and lean_proofsteps train 12,699,102 bytes (12.7 MB zst) both confirmed via blobs=true; validation and ",
   "pre-Oct-2023 snapshot (repo last modified 2023-10-25), so mostly Lean 3 plus early Lean 4",
   "unknown (cardData license absent; matches candidate)",
   "Fetch only the two Lean shards (about 115 MB compressed) for GitHub Lean files and LeanDojo-derived proofsteps; do not pull the 120 GB repo.",
   "Medium. Lean 3 era GitHub files may include miniF2F Lean 3 solutions and ProofNet-adjacent material. Lean 3 syntax is also incompatible with our Lean 4 target, so filter hard.",
   "API 200 (HF API); datasets-server 501 (script-based, no viewer rows) 2026-10-14; 901a9273a770e9d4138c5ddd91802f9c5c6cdc4b",
   "https://arxiv.org/abs/2310.10631",
   "Med",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus",
   "hoskinson-center/proof-pile (v1)",
   "hoskinson-center/proof-pile",
   "HF-dataset",
   "corpus",
   "First-generation Proof-Pile (arXiv math, formal libraries incl. Lean 3 mathlib, Isabelle, Coq, etc.).",
   "datasets-server (partial=true): train 270,000 rows, 5.01 GB num_bytes; validation 46,251; test 46,251; dataset_size 6.15 GB. usedStorage 42,155,055,536 bytes (~",
   "mostly Lean 3 (unverified)",
   "apache-2.0 (cardData list form ['apache-2.0']; matches candi",
   "Probably skip; superseded by proof-pile-2 and has no Lean-named file in the sibling list I checked.",
   "Medium (Lean 3 miniF2F-adjacent content possible).",
   "API 200 (HF API and datasets-server info) 2026-10-14; 490b980249446f2f3bd2df3a8cf085d0f2de240a",
   "https://arxiv.org/abs/2302.12433",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus",
   "sean-lamont/leandojo_traces",
   "sean-lamont/leandojo_traces",
   "HF-dataset",
   "traces",
   "Zipped LeanDojo search traces and models (goal_model, htps, bestfs traces, original_train, etc.).",
   "usedStorage 16,136,085,024 bytes (~16.1 GB, matches candidate); 8 zip files (goal_model, goal_model_higher_tp, htps, original_no_retriever_bestfs_60_acc, origin",
   "unknown",
   "unknown (cardData license absent; matches candidate)",
   "Not raw Lean text. Mostly search artefacts; low priority for S1, possibly relevant to S2 traces.",
   "Medium (includes test-split search logs for LeanDojo benchmark; not miniF2F).",
   "API 200 (HF API); datasets-server error: BadZipFile (zip archives, no parquet config) 2026-10-14; 250c349a2f52e02cb4423cb8c467bfa2df57e025",
   "workflow wf_c3f2e9d0-871",
   "Med",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus",
   "tinyproof0/Lean4-4.14-LeanUniverse",
   "tinyproof0/Lean4-4.14-LeanUniverse",
   "HF-dataset",
   "corpus",
   "A LeanUniverse-style crawl of many Lean 4 GitHub repos built against Lean 4.14 (directories licenses/, novel_premises/, random/ seen in the tree).",
   "rows unknown: datasets-server config-info FAILED (no info returned). usedStorage 962,332504 bytes (~962 MB, matches candidate). corpus.jsonl size not reported i",
   "4.14 (stated in name)",
   "None in cardData (matches candidate). Per-repo licenses/ dir",
   "Compiled, multi-repo Lean 4 code beyond Mathlib; license folder allows filtering by repo license.",
   "Medium. GitHub crawl; may contain miniF2F/ProofNet solutions in some repos.",
   "API 200 2026-10-14; 20af9e1aab60a2670eeec1e2689282f78662cfa4",
   "workflow wf_c3f2e9d0-871",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs",
   "deepseek-ai/DeepSeek-Prover-V1",
   "deepseek-ai/DeepSeek-Prover-V1",
   "HF-dataset",
   "corpus",
   "Synthetic Lean 4 theorem plus proof pairs (name, split, formal_statement, goal, header, formal_proof).",
   "27,503 rows (train, datasets-server); single dataset.jsonl; repo usedStorage 51,571,902 bytes",
   "Lean 4 (v4.9.0-era likely, unverified)",
   "other (cardData.license=other; LICENSE file present, text no",
   "Small model-generated Lean; use only as filler, not as the core of mid-training.",
   "Medium; the DeepSeek-Prover team reported decontaminating against miniF2F, but verify yourself.",
   "API 200 (api) / 200 (datasets-server) 2026-10-14; 1ee889f608fb12ba3596757ee91a60acd663ea81",
   "https://arxiv.org/abs/2405.14333",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs",
   "Goedel-LM/Lean-workbook-proofs",
   "Goedel-LM/Lean-workbook-proofs",
   "HF-dataset",
   "corpus",
   "Full Lean 4 proofs for Lean Workbook problems (problem_id, full_proof), produced by Goedel-Prover.",
   "29,750 rows (train split, datasets-server); repo usedStorage 16,165,753 bytes; parquet bytes unverified",
   "Lean 4 (version unverified)",
   "unknown (cardData.license=None; no license tag; repo contain",
   "Same lineage as our base model data; including it would replicate what Goedel-Prover-V2-8B already saw, so use with care for novelty.",
   "Medium-High; Lean Workbook overlap, model-generated proofs.",
   "API 200 (api) / 200 (datasets-server) 2026-10-14; b731852af8d8ab11498fda27bce9020738c01c59",
   "https://arxiv.org/abs/2502.07640",
   "Med",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs",
   "kfdong/STP_Lean_SFT",
   "kfdong/STP_Lean_SFT",
   "HF-dataset",
   "sft-data",
   "Self-play (STP) generated Lean 4 prompt/target pairs, plus an eval split.",
   "235,430 train + 4,096 eval rows (datasets-server); repo usedStorage 871,197,069 bytes",
   "Lean 4 (version unverified)",
   "unknown (cardData.license=None; no license tag; no LICENSE f",
   "More S2 than S1; noted here because it is a source of verified Lean proofs.",
   "Medium-High; built on Lean Workbook/miniF2F-adjacent seed sets.",
   "API 200 (api) / 200 (datasets-server) 2026-10-14; 7166a1964c466af947ae804a09857d69498b1b99",
   "https://arxiv.org/abs/2502.00212",
   "Med",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs; S2b SFT agent/correction",
   "FrenzyMath/Herald_proofs",
   "FrenzyMath/Herald_proofs",
   "HF-dataset",
   "corpus",
   "Herald: Lean 4 formal theorem, informal theorem, formal proof, informal proof, commented proof, header, from Mathlib-derived autoformalisation.",
   "44,553 rows; num_bytes 161,927,210 (~162 MB); download_size 57,908,255; usedStorage 115,217,715",
   "Lean 4 (Mathlib v4.16-era per sibling dataset name; unverified for this repo)",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Parallel NL-Lean text, usable in mid-training or as SFT filler; the commented_proof field is useful comment-rich Lean.",
   "Low-Medium. Derived from Mathlib; informal statements are LLM-translated. Check for overlap with competition-style statements.",
   "API 200 2026-10-14; 2d31b459df74d730bff85682de237d894b4eecfa",
   "https://arxiv.org/abs/2410.10878",
   "High",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs; S2b SFT agent/correction; S3 RL prompts",
   "Goedel-LM/Goedel-Pset-v1",
   "Goedel-LM/Goedel-Pset-v1",
   "HF-dataset",
   "corpus",
   "Large pool of autoformalised Lean 4 problem statements (problem_id, formal_statement, informal_statement); statements only.",
   "1,732,594 rows (train, datasets-server); 4 parquet shards; repo usedStorage 827,628,391 bytes",
   "Lean 4 (version unverified)",
   "unknown (cardData.license=None; no license tag; repo contain",
   "Statement text only. Better for S3 prompts than S1 corpus.",
   "Medium. Autoformalised from NuminaMath and others; possible overlap with miniF2F-style items.",
   "API 200 (api) / 200 (datasets-server) 2026-10-14; 79fc4fbe87b4c694a941bc63ad92dc38d39d41db",
   "https://arxiv.org/abs/2502.07640",
   "Med",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs; S3 RL prompts",
   "AI-MO/NuminaMath-LEAN",
   "AI-MO/NuminaMath-LEAN",
   "HF-dataset",
   "corpus",
   "Competition problems with Lean 4 formal statements (uuid, problem, question_type, answer, author, formal_statement); proof columns not confirmed.",
   "104,155 rows (train, datasets-server); repo usedStorage 157,131,057 bytes",
   "Lean 4 (version unverified)",
   "apache-2.0 (cardData and license tag)",
   "Mostly an S3 prompt source; small for S1.",
   "Medium; competition-derived.",
   "API 200 (api) / 200 (datasets-server) 2026-10-14; 51fa67f1f647ae1ecd81eef9f19306aa8a7b3a94",
   "https://arxiv.org/abs/2504.11354",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs; S3 RL prompts",
   "internlm/Lean-Github",
   "internlm/Lean-Github",
   "HF-dataset",
   "corpus",
   "Tactic-level records mined from GitHub Lean 4 repos: url, commit, file_path, full_name, start, end, tactic, state_before.",
   "218,866 rows (datasets-server); num_bytes 436,501,379 (~437 MB uncompressed, matches candidate); on-disk usedStorage 84,088,551 bytes (~84 MB parquet); 1 file l",
   "mixed, per-repo commits",
   "apache-2.0 (cardData; license tag matches)",
   "Broad non-Mathlib Lean coverage (state-before and tactic pairs), good as a proof-step supplement. Note it holds tactic pairs, not whole files.",
   "Medium. GitHub crawl may include repos that formalise miniF2F/ProofNet solutions. Decontaminate by statement match.",
   "API 200 (HF API and datasets-server info) 2026-10-14; 22220a1073f959ec6fe1cee596965a8275bcc550",
   "https://arxiv.org/abs/2407.17227",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S2a SFT proofs; S3 RL prompts",
   "internlm/Lean-Workbook",
   "internlm/Lean-Workbook",
   "HF-dataset",
   "corpus",
   "Lean 4 competition-style problems with proven tactic-state data (id, status, tactic, state_before, state_after, natural_language_statement).",
   "25,214 rows (train split, datasets-server confirmed); num_bytes 16,597,556 (~17 MB); download_size 4,600,174 (~4.6 MB); usedStorage 103,685,829 (~104 MB)",
   "Lean 4 (version unverified)",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Small as a corpus, better in S3 or S4; include only lightly in mid-training.",
   "Medium-High. Problems include AoPS and competition-adjacent statements; overlap with miniF2F is possible. Run exact and fuzzy match.",
   "API 200 2026-10-14; 2e066e310b2c6d2c27616927ae131f82901c8f1c",
   "https://arxiv.org/abs/2406.03847",
   "High",
   "Decontaminate first"
  ],
  [
   "S1 Mid-train corpus; S3 RL prompts",
   "FrenzyMath/Herald_statements",
   "FrenzyMath/Herald_statements",
   "HF-dataset",
   "corpus",
   "Herald formal and informal theorem statements (without proofs).",
   "579,883 rows (train split, previously unverified); num_bytes 310,480,955 (~310 MB); download_size 133,063,027 (~133 MB); usedStorage 133,063,027",
   "unknown",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Statement-only text; helpful for statement-writing proficiency but low proof density.",
   "Low-Medium.",
   "API 200 2026-10-14; d8d849682ecfba92dda26026a97e10a662187f4c",
   "https://arxiv.org/abs/2410.10878",
   "Med",
   "Candidate"
  ],
  [
   "S1 Mid-train corpus; S3 RL prompts",
   "phanerozoic/Lean4-FormalConjectures",
   "phanerozoic/Lean4-FormalConjectures",
   "HF-dataset",
   "corpus",
   "Declarations from Google DeepMind formal-conjectures repo (open and solved problem statements in Lean).",
   "4557 rows (datasets-server info, train split; candidate said unverified); dataset_size 1809583 B; download_size 836301 B; usedStorage 15331207 B (~15 MB)",
   "unknown",
   "apache-2.0 (cardData)",
   "Small, high-value source of prove-or-disprove style statements; also useful later for S3 prompts.",
   "Medium. Repo covers some Erdos/competition-adjacent problems and may overlap PutnamBench/miniF2F-style statements. Check before use.",
   "API 200 2026-10-14; sha 14a2974933546cd0b068ca6825643f6e7e4db9f6; card says sour",
   "workflow wf_c3f2e9d0-871",
   "Med",
   "Decontaminate first"
  ],
  [
   "S2a SFT proofs",
   "DeepSeek-Prover-V2 generation data (mirror)",
   "Cartinoe5930/DeepSeek-Prover-V2-dataset-new",
   "HF-dataset",
   "traces",
   "'messages' chat column; third-party reconstruction of DeepSeek-Prover-V2-style data",
   "66,722 train rows (datasets-server); 21,359,902 bytes total storage",
   "unknown",
   "unverified (no license in API tags or cardData; candidate 'u",
   "Possible CoT-style proof traces. Provenance is unclear; treat as low priority.",
   "Unverified. Provenance unknown.",
   "API api=200, datasets-server info=200 2026-10-14; a8c903b3ee9593df42ad7427c135b77cbb4bd3ef",
   "https://arxiv.org/abs/2504.21801",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "FineLeanCorpus",
   "m-a-p/FineLeanCorpus",
   "HF-dataset",
   "corpus",
   "NL statement paired with a Lean 4 formalization: id, statement, lean_code, compile_result, eval_reasons, difficulty, difficulty_rationale, domain. Card says each entry passes the Lean compiler plus CriticLean semantic check; 79.19% accuracy from manual sampling.",
   "509,358 rows (datasets-server train); dataset_size 1.51 GB; usedStorage 14.69 GB across 26 files",
   "unknown",
   "unknown (cardData license is None and no license: tag)",
   "Statement source for expert iteration, and for stratifying by difficulty. The lean_code is a formalization, so it may not contain proofs; check before using as SFT.",
   "Unverified. Draws on many public competition sources; check against miniF2F.",
   "API 200 2026-10-14; e32e721f08698ffc2821af8b8248000b0464318e",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "FormalMATH SFT (classic)",
   "formalmathatepfl/sft_classic",
   "HF-dataset",
   "sft-data",
   "uuid, data_source, question, answer, proof_plan, valid, proof_repair, lean_code",
   "train 2,678,749; test 623,741; sft_replay_classic 161,030 rows (datasets-server); 13,705,211,170 bytes total storage; download 5,487,250,864 bytes; 46 files",
   "unknown",
   "unverified (no license in API tags or cardData; candidate 'u",
   "Large pool with a 'valid' flag, so we can filter to verified proofs. Provenance of this EPFL-related set is unclear from the metadata.",
   "Unverified; the test split might overlap benchmarks.",
   "API api=200, datasets-server info=200 2026-10-14; 21027890d0208fdc3881bb79b969e7bdf01dabbf",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Goedel workbook proofs migrated to Lean 4.27",
   "banach1729/goedel-workbook-lean427",
   "HF-dataset",
   "sft-data",
   "60,341 traced (state, tactic) pairs from 29,750 Goedel Workbook proofs, migrated to Lean 4.27.0 / Mathlib v4.27.0. 28,016 (94.1%) of the proofs compile on 4.27.",
   "unverified. datasets-server returned a failed config-info, so no row count is confirmed. API siblings: 10,006 entries, all Lean source files under GoedelProofs/",
   "4.27.0 / Mathlib v4.27.0 (card)",
   "apache-2.0 (cardData.license confirmed)",
   "Only useful if we move to a recent Lean. Lean 4.27 does not match the miniF2F Lean 4.9.0-rc1 header, so I would not use it for this eval.",
   "Same as the upstream Lean Workbook proofs.",
   "API 200 (api) / datasets-server config-info failed 2026-10-14; sha 4049a8c4c2c7ae05d41207dccba1e9bf8afeb664",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "iiis-lean math formal corpus",
   "iiis-lean/lean-math-formal-corpus",
   "HF-dataset",
   "corpus",
   "2,877 problems per config with uid, dataset, source_id, source_origin, source_lean_version, nl_problem. Separate configs for Lean v4.9.0, v4.15.0 and v4.27.0.",
   "8,631 rows",
   "v4.9.0, v4.15.0, v4.27.0 configs",
   "unknown (no cardData license)",
   "Useful because it ships a v4.9.0 config, close to our 4.9.0-rc1 header. Check which source benchmarks it contains (may include miniF2F or ProofNet).",
   "Unchecked; only Mathlib-derived set found with a v4.9.0 config (matches miniF2F header) - decontaminate by statement sha",
   "API 200 2026-10-14; sha fc3e0ed3; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Lean Workbook (full statement list)",
   "pkuAI4M/LeanWorkbook",
   "HF-dataset",
   "corpus",
   "formal_statement plus natural_language_statement",
   "140,214 rows (datasets-server num_examples, train split confirmed); 46,962,716 bytes uncompressed; 22,033,462 bytes download; 3 files",
   "unknown",
   "unknown (cardData.license None, no license tag; matches cand",
   "Larger statement pool for RL prompts; no proofs.",
   "Same as Lean Workbook.",
   "API 200 (api) / 200 (datasets-server info) 2026-10-14; sha b7b59dd09fb6e771290ed616a7037bc62a16842a",
   "https://arxiv.org/abs/2406.03847",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Lean-STaR plus",
   "ScalableMath/Lean-STaR-plus",
   "HF-dataset",
   "sft-data",
   "input/output pairs: tactic-step prediction with interleaved natural-language thoughts, per the Lean-STaR paper",
   "61,082 rows (datasets-server train); 62.2 MB download; single file STaR-generated-train.json",
   "unknown (the paper builds on LeanDojo Lean 4; unverified)",
   "unknown (cardData license is None and no license: tag)",
   "CoT-flavoured tactic data, useful if we want rationale-style outputs.",
   "Unverified; derived from the LeanDojo Mathlib benchmark.",
   "API 200 2026-10-14; 2a6b3d5b92c09b498e80fab687bb7b2d2cfb9651",
   "https://arxiv.org/abs/2407.10040",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Lean4 subgoal completions",
   "connorolson/lean4-subgoal-completions",
   "HF-dataset",
   "sft-data",
   "Subgoal completion pairs. Schema and size not retrieved.",
   "unknown",
   "unknown",
   "cc-by-4.0",
   "Auxiliary SFT; verify content.",
   "Unknown.",
   "API 200 2026-10-14; sha 4f04d3e3; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "LeanDojo Benchmark 4 (random split, mirror)",
   "cat-searcher/leandojo-benchmark-4-random",
   "HF-dataset",
   "sft-data",
   "Mathlib4 (state, tactic, target_state) with url, commit, file_path, split by random",
   "250,814 train / 4,260 validation / 4,506 test (datasets-server); 48.8 MB download; 5 parquet files",
   "unknown (commit field per row; unverified)",
   "unknown (cardData license is None and no license: tag; LeanD",
   "Tactic-level pretraining or SFT on Mathlib. Third-party mirror of LeanDojo Benchmark 4, since the official kaiyuy/* ids I tried returned errors.",
   "Low for miniF2F, but high for any Mathlib-derived eval. Not a decontamination risk for competition problems.",
   "API 200 2026-10-14; 8a88a90126dd900e33039562766d6b9b60c66fbf",
   "https://arxiv.org/abs/2306.15626",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "LeanDojo data hammer",
   "hanwenzhu/leandojo_data_hammer",
   "HF-dataset",
   "sft-data",
   "LeanDojo-derived data used for hammer-style training. Schema not retrieved.",
   "unknown (17 files)",
   "unknown",
   "unknown (no cardData license)",
   "Auxiliary tactic data; low priority.",
   "Unknown (LeanDojo Mathlib base).",
   "API 200 2026-10-14; sha 948836e9; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "LeanDojo-filtered",
   "LLParallax/LeanDojo-filtered",
   "HF-dataset",
   "sft-data",
   "58,992 theorems with url, commit, file_path, full_name, start, end, traced_tactics.",
   "58,992 rows",
   "unknown (commit field per row)",
   "unknown (no cardData license)",
   "Commit-pinned traced tactics; commit and file path fields allow rebuilding against a chosen Mathlib commit.",
   "Low to moderate.",
   "API 200 2026-10-14; sha 771a7056; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Mathlib subproofs (chasenorman)",
   "chasenorman/subproofs-mathlib-v4.30.0",
   "HF-dataset",
   "sft-data",
   "433,293 subproof records with type, tactic, removals, name, kind, goal. A rollout variant (chasenorman/rollout-subproofs-mathlib-v4.30.0, 4,139 rows) adds result fields.",
   "433,293 rows",
   "Mathlib v4.30.0 (stated in id; v4.29.0 and others also exist)",
   "unknown (no cardData license)",
   "Subgoal-level SFT or curriculum data. Version v4.30.0 is much newer than our header, so needs re-checking.",
   "Low.",
   "API 200 2026-10-14; sha 1451d85e; gated=False",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "Nemotron-Math-Proofs-v1 (NVIDIA)",
   "nvidia/Nemotron-Math-Proofs-v1",
   "HF-dataset",
   "sft-data",
   "Lean problems with formal_statement, lean_header, source and url fields. Per search snippet, proofs are model-generated and compiler-checked, with error-feedback self-correction. Proof columns are not in the first 8 listed fields.",
   "243,314 rows",
   "unknown (lean_header column present, not read)",
   "cc-by-sa-4.0",
   "SFT whole-proof data and a source of statements for the RL prompt pool. lean_header lets us check Lean version compatibility.",
   "Likely. Sources include public statement sets; not measured.",
   "API 200 2026-10-14; sha 97229c59; gated=False",
   "unverified",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "STP_Lean (self-play theorem prover)",
   "kfdong/STP_Lean",
   "HF-dataset",
   "sft-data",
   "prompt, target, weight, iteration, tag. Self-play conjecture-and-proof training examples.",
   "1,738,754 rows (datasets-server num_examples, train split confirmed); 3,270,624,094 bytes uncompressed; 2,429,396,930 bytes usedStorage; 9 files",
   "unknown (STP used Lean 4 / Mathlib; verify)",
   "mit (cardData.license and license tag confirmed)",
   "Large pool of proofs found by self-play. Weights allow subsampling.",
   "Unverified. Prompts mix Lean Workbook and LeanDojo-style statements; check miniF2F overlap.",
   "API 200 (api) / 200 (datasets-server info) 2026-10-14; sha 30a52fa0b690a272c76ac42765345e64321d48d0",
   "https://arxiv.org/abs/2502.00212",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs",
   "PutnamBench Lean 4 (decontamination reference)",
   "ChristianZ97/PutnamBench-lean4",
   "HF-dataset",
   "benchmark",
   "problem_name, formal_statement, informal_statement, informal_solution, tags, split",
   "train 672 rows (datasets-server); 239,316 bytes total storage",
   "unknown",
   "apache-2.0",
   "Overlap check for SFT data. Not training data.",
   "Eval set. Do not train on it.",
   "API api=200, datasets-server info=200 2026-10-14; f753021f5fd7e40da05e137c0d0cb8624790f227",
   "https://arxiv.org/abs/2407.11214",
   "Med",
   "Eval only"
  ],
  [
   "S2a SFT proofs; S2b SFT agent/correction",
   "Goedel-Prover-V2 SFT dataset",
   "Goedel-LM/SFT_dataset_v2",
   "HF-dataset",
   "sft-data",
   "Single 'messages' column (chat format) SFT data used for Goedel-Prover-V2; the card gives no field-level description or per-source breakdown",
   "1,745,010 rows (datasets-server num_examples, train split confirmed); 26,392,353,371 bytes uncompressed; 6,106,217,900 bytes download; 56 files",
   "unknown (card silent)",
   "apache-2.0 (cardData.license and license tag confirmed)",
   "Highest-priority candidate. It is the data behind our base model, so it is a sanity reference and a replay-mix component to limit forgetting. Inspect first-rows before use.",
   "Unverified. Card has no decontamination statement; check against miniF2F-test before mixing.",
   "API 200 (api) / 200 (datasets-server info) 2026-10-14; sha da71b450f771880ea314a8860532ee26bf494753",
   "https://arxiv.org/abs/2508.03613",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs; S2b SFT agent/correction",
   "Leanabell-Prover SFT train data",
   "stoney0062/Leanabell-Prover-Traindata-SFT",
   "HF-dataset",
   "sft-data",
   "prompt and output pairs for Leanabell-Prover SFT",
   "218,988 rows (datasets-server train); 376.6 MB download; single file traindata_sft_without_cot.jsonl",
   "unknown",
   "apache-2.0 (cardData and tag)",
   "Extra SFT pool, with possible long-CoT-style outputs; inspect first-rows.",
   "Unverified.",
   "API 200 2026-10-14; 1f1faf82920cb8590dc5c8c091dd46ba91a133bf",
   "https://arxiv.org/abs/2504.06122",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs; S2b SFT agent/correction",
   "Leanabell-Prover-V2 cold-start data",
   "stoney0062/Leanabell-Prover-V2-Coldstart-DS",
   "HF-dataset",
   "sft-data",
   "'conversations' column, cold-start chats for Leanabell-Prover-V2",
   "5,638 rows (datasets-server train); 55.7 MB download; single file coldstart_data.json",
   "unknown",
   "apache-2.0 (cardData and tag)",
   "Small cold-start set. May include multi-turn repair/feedback format, useful for the later multi-turn stage (inspect first).",
   "Unverified.",
   "API 200 2026-10-14; d6a6d5048521d4b2dab60479ac3a4859fb87b3dd",
   "https://arxiv.org/abs/2507.08649",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs; S2b SFT agent/correction; S3 RL prompts",
   "Kimina-Prover Promptset",
   "AI-MO/Kimina-Prover-Promptset",
   "HF-dataset",
   "rl-prompts",
   "statement_id, natural_language, formal_statement, source, name. No proofs.",
   "24,418 rows (datasets-server num_examples, train split confirmed); 15,135,350 bytes uncompressed; 4,360,794 bytes download; 3 files",
   "unknown",
   "unknown (no cardData license and no license tag; matches can",
   "Prompt pool for generating our own verified proofs.",
   "Unverified. Likely includes problems derived from miniF2F-like sources.",
   "API 200 (api) / 200 (datasets-server info) / 200 (first-rows) 2026-10-14; sha 3009c548d90160d0f5e963d72238610c6732f812",
   "https://arxiv.org/abs/2504.11354",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs; S3 RL prompts",
   "STP_Lean_0320",
   "kfdong/STP_Lean_0320",
   "HF-dataset",
   "sft-data",
   "Same schema as STP_Lean (prompt, target, weight, iteration, tag), a later and larger snapshot",
   "3,262,558 rows (datasets-server train split); ~1.74 GB download, 5.77 GB dataset_size; 10 parquet shards",
   "unknown",
   "mit (cardData and license:mit tag)",
   "Alternative to STP_Lean; use whichever is the latest snapshot and deduplicate against the other.",
   "Unverified, as for STP_Lean.",
   "API 200 2026-10-14; 8759699c72a373a49497cde1341a4a94c3845b2f",
   "https://arxiv.org/abs/2502.00212",
   "Med",
   "Candidate"
  ],
  [
   "S2a SFT proofs; S4 Verifier/infra; S6 Eval (held-out)",
   "miniF2F Lean 4 (eval, for decontamination)",
   "cat-searcher/minif2f-lean4",
   "HF-dataset",
   "benchmark",
   "miniF2F valid and test, 244 each, with header, formal_statement, informal_stmt, informal_proof",
   "validation 244; test 244 rows (datasets-server); 332,956 bytes total storage",
   "unknown (the Goedel header in our project is Lean 4.9.0-rc1)",
   "mit",
   "Reference for n-gram/AST overlap checks against every SFT dataset above. Not training data.",
   "This is the eval itself. Keep miniF2F-test out of training, and use valid only for dev.",
   "API api=200, datasets-server info=200 2026-10-14; 70a1249ce240667f6bcdd1ccd62f847f0e065d57",
   "https://arxiv.org/abs/2109.00110",
   "Med",
   "Eval only"
  ],
  [
   "S2a SFT proofs; S6 Eval (held-out)",
   "ProofNet (original)",
   "hoskinson-center/proofnet",
   "HF-dataset",
   "benchmark",
   "Undergraduate-level formal statements: nl_statement, nl_proof, formal_statement, src_header",
   "test 186; validation 185 rows (datasets-server, config plain_text); 12,243,856 bytes total storage",
   "Lean 3 original (the formal_statement is Lean 3; Lean 4 ports exist as third-party mirrors such as UDACA/proofnet-lean4; unverified)",
   "mit",
   "Overlap check for SFT data. Not training data.",
   "Eval set. Do not train on it.",
   "API api=200, datasets-server info=200 2026-10-14; cb8e75614830035a37f3a2a11de5e625eaf0bc31",
   "https://arxiv.org/abs/2302.12433",
   "High",
   "Eval only"
  ],
  [
   "S2b SFT agent/correction",
   "Douvras Lean proof-repair corpus",
   "dougdotcon/douvras-lean-proof-repair",
   "HF-dataset",
   "sft-data",
   "36 rows total (24/6/6) with failed_code, compiler_error, candidate_correction. The card says the snippets are synthetic and were NOT executed in the compiler.",
   "36 rows total: train 24, validation 6, test 6 (datasets-server); dataset_size 13,385 bytes; usedStorage 26,678 bytes",
   "unknown (mathlib_version column exists; values unverified)",
   "cc-by-4.0 (cardData matches candidate)",
   "Not usable as training data (unverified, tiny). Schema only.",
   "Low (synthetic).",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; 827461ee4da59ee6ab79d8181de513da3c29eaf6",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "Erdos 741 II Lean 4 Opus traces",
   "vincentoh/erdos741ii-lean4-opus-traces",
   "HF-dataset",
   "traces",
   "39 Claude Opus agent sessions with messages, reward, n_tool_calls, n_oracle_runs, last_oracle_score, ever_scored_1.",
   "39 rows in train split; usedStorage 776,102 bytes (under 0.01 GB, confirmed); dataset_size 802,068 bytes",
   "unknown",
   "apache-2.0 (cardData matches candidate)",
   "Tiny; useful as a format check only.",
   "Low.",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; 2a87ad63a73585e6c9610799842035d9a0f0a209",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "Lean Copilot",
   "https://github.com/lean-dojo/LeanCopilot",
   "GitHub",
   "tool",
   "LLM tactic suggestion and proof search inside Lean 4. 1,333 stars, pushed 2026-10-13.",
   "1402 KB (repo size per GitHub API)",
   "tracks recent Lean 4 releases (unverified)",
   "MIT (SPDX MIT, 'MIT License')",
   "Reference for tool-call integration; no trace dataset.",
   "n/a",
   "API 200 2026-10-14; default_branch=main; commit sha not fetched",
   "https://arxiv.org/abs/2404.12534",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "LeanAgent",
   "https://github.com/lean-dojo/LeanAgent",
   "GitHub",
   "tool",
   "Lifelong learning agent for Lean theorem proving, 87 stars, last push 2025-06-13.",
   "15558 KB (repo size per GitHub API)",
   "unknown",
   "unverified (API license field is null; no SPDX id returned)",
   "Reference only; a ReProver/LeanDojo-style tactic-level prover and not a code-agent trace source.",
   "n/a",
   "API 200 2026-10-14; default_branch=main; commit sha not fetched",
   "https://arxiv.org/abs/2410.06209",
   "Med",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "Nemotron-Cascade SFT SWE",
   "nvidia/Nemotron-Cascade-SFT-SWE",
   "HF-dataset",
   "sft-data",
   "SWE-focused SFT data from NVIDIA Nemotron-Cascade; size category 100K-1M rows.",
   "162,262 rows (split train across swe_localization.jsonl, swe_repair.jsonl, swe_testgen.jsonl); dataset_size 5.01 GB; download size null; repo usedStorage 8.53 G",
   "n/a",
   "cc-by-4.0 (cardData and tag agree)",
   "Large generic code-agent SFT data. Check format (agentless vs agentic) before use.",
   "None for Lean.",
   "API 200 (api, datasets-server info) 2026-10-14; b4ef73ff9bbcfc33d1ec7a48e53017f9ce7af7a3",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "Numina-Lean-Agent",
   "https://github.com/project-numina/numina-lean-agent",
   "GitHub",
   "tool",
   "Agentic Lean prover scaffold from Project Numina (279 stars, pushed 2026-07-08). I confirmed only that the repo exists via the API. I did not read the README, so tool details and any released traces are unverified.",
   "4376 KB (repo size per GitHub API)",
   "unknown",
   "unverified (API license field is null; no SPDX id returned)",
   "Candidate harness for producing Lean agent trajectories with our own model, with the Lean build as reward.",
   "n/a (tool).",
   "API 200 2026-10-14; default_branch=main; commit sha not fetched (repos endpoint ",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "RRMA Lean 4 agent traces",
   "vincentoh/rrma-lean4-agent-traces",
   "HF-dataset",
   "traces",
   "416 multi-agent Lean 4 proof-search sessions on Erdos #741(ii) and #125. Columns: messages, reward, n_turns, n_tool_calls, n_oracle_runs, last_oracle_score, rung. Generated by Claude Opus/Haiku/Sonnet agents with a build oracle. Card says 115 of 416 have reward=1.0 (v2 relabel).",
   "416 rows in train split (datasets-server); usedStorage 30,843,511 bytes (~0.03 GB), dataset_size 35,200,490 bytes. Candidate's 'about 0.01 GB' is contradicted (",
   "unknown",
   "apache-2.0 (cardData matches candidate)",
   "Format reference and a small seed for the Lean code-agent half (tool-call and oracle-feedback loop). Too small and too narrow for training alone. Use only reward=1.0 rows.",
   "Low for miniF2F/PutnamBench (Erdos problems). Teacher outputs are Claude-generated, so check provider terms for distillation.",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; d91b6e45486332cd53eb1ba072852e4b00219566",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "SWE-Gym OpenHands SFT trajectories",
   "SWE-Gym/OpenHands-SFT-Trajectories",
   "HF-dataset",
   "traces",
   "491 successful OpenHands trajectories on SWE-Gym tasks, as chat messages.",
   "491 rows (split train.success.oss); dataset_size 31.8 MB; download 10.4 MB; repo usedStorage 10.4 MB (about 0.01 GB, matches)",
   "n/a",
   "mit (cardData and tag agree)",
   "Small high-quality general code-agent SFT seed.",
   "None for Lean benchmarks. Possible overlap with SWE-bench Verified repos.",
   "API 200 (api, datasets-server info) 2026-10-14; 4aaa5a4a4b5861f4799d2336908760c190ac3b17",
   "https://arxiv.org/abs/2412.21139",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "SWE-Hero OpenHands trajectories (NVIDIA)",
   "nvidia/SWE-Hero-openhands-trajectories",
   "HF-dataset",
   "traces",
   "OpenHands SWE trajectories released by NVIDIA (2026-05). Row count not confirmed.",
   "34,269 rows (split train, 14 parquet shards); dataset_size 8.49 GB; download 2.40 GB; repo usedStorage 14.67 GB",
   "n/a",
   "cc-by-4.0 (cardData and tag agree)",
   "Alternative generic agent traces under cc-by-4.0.",
   "None for Lean.",
   "API 200 (api, datasets-server info) 2026-10-14; 150bc119e52c647216fce285fd801f16b6fd745b",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "SWE-rebench OpenHands trajectories",
   "nebius/SWE-rebench-openhands-trajectories",
   "HF-dataset",
   "traces",
   "67,074 OpenHands trajectories with resolved flags, tools, model_patch and generated-test pass info.",
   "67,074 rows (split train); dataset_size 16.6 GB; download 2.08 GB; repo usedStorage 2.08 GB (about 2.1 GB matches)",
   "n/a",
   "cc-by-4.0 (cardData and tag agree; LICENSE file present)",
   "Large general code-agent SFT pool; filter resolved.",
   "None for Lean.",
   "API 200 (api, datasets-server info) 2026-10-14; 35455389ab51bf5e2306bfd436ef72d0f98bf882",
   "https://arxiv.org/abs/2505.20411",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "SWE-smith trajectories",
   "SWE-bench/SWE-smith-trajectories",
   "HF-dataset",
   "traces",
   "Agent trajectories on SWE-smith tasks with messages, resolved flag, patch; three tool-format variants (tool 24,100 / xml 26,076 / ticks 25,826 rows).",
   "76,002 rows across splits tool 24,100 + xml 26,076 + ticks 25,826; dataset_size 9.30 GB; download 3.17 GB; repo usedStorage 8.78 GB",
   "n/a",
   "mit (cardData and tag agree)",
   "Largest MIT-licensed generic agent trace source here; filter resolved=true and pick the tool-call format.",
   "None for Lean; Python repos.",
   "API 200 (api, datasets-server info) 2026-10-14; 08e109b4a59eaeebf80e4675cd125d42e7ac99a4",
   "https://arxiv.org/abs/2504.21798",
   "High",
   "Candidate"
  ],
  [
   "S2b SFT agent/correction",
   "Goedel-Prover-V2 RL dataset (lean4_revision / lean4_completion subsets)",
   "Goedel-LM/RL_dataset_V2",
   "HF-dataset",
   "rl-prompts",
   "98,632 rows with 'subset' and 'messages'. First rows show subsets 'lean4_revision' (self-correction format) and 'lean4_completion'. Statements include OMR and lean_workbook.",
   "train 98,632 rows (datasets-server); download 336,508,509 bytes (about 0.34 GB, confirmed); 1,049,807,386 bytes on disk",
   "unknown",
   "apache-2.0",
   "The revision subset is the closest released analogue of compiler-feedback correction traces. Also usable as a source of single-theorem RL prompts after the rarely-solved-but-not-never filter.",
   "Medium. Same statement sources as SFT_dataset_v2; no miniF2F decontamination statement confirmed.",
   "API api=200, datasets-server info=200 2026-10-14; 5c99dce377f387bd271d726f205e69380b792a9d",
   "https://arxiv.org/abs/2508.03613",
   "Med",
   "Decontaminate first"
  ],
  [
   "S2b SFT agent/correction",
   "Lean proof-or-refute 300",
   "xlr8harder/lean-proof-or-refute-300",
   "HF-dataset",
   "benchmark",
   "300 test prompts with answer polarity, certificate, mutation_kind (statements mutated so some are false). Matches the prove-or-disprove task shape.",
   "300 rows in test split; dataset_size 1,138,538 bytes; usedStorage 289,378 bytes (under 0.01 GB, confirmed)",
   "unknown",
   "mit (cardData matches candidate)",
   "Dev or eval for the prove-or-disprove 30% slice; its mutation recipe can be copied to make training problems. 300 rows is too small to train on.",
   "Has a 'development_exposure' column; origin of statements unverified. Check overlap with miniF2F before mixing.",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; 7b6990f99580820c553573bfbd18132f3a28228f",
   "unknown",
   "Med",
   "Eval only"
  ],
  [
   "S2b SFT agent/correction",
   "LeanPolish proof-compression traces",
   "leanpolish-anon/lean-proof-compression",
   "HF-dataset",
   "traces",
   "Configs goedel, mathlib, minif2f, putnam_bench, putnam_verified, putnam2025: proof-shortening edits with failed_attempts and failed_tactics, split into training/rejected. goedel config has 20,822 training and 28,525 rejected rows.",
   "goedel config: training 20,822 + rejected 28,525 = 49,347 rows (candidate's 49,347 confirmed in total, but training split alone is 20,822). mathlib config: trai",
   "unknown",
   "apache-2.0 (cardData matches candidate)",
   "Verified edit/rejection pairs for proof-rewrite behaviour. Use only the goedel and mathlib configs.",
   "HIGH for minif2f and putnam_bench configs, which are built from those benchmarks. Exclude them.",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; 64422193ada0449e2a60d4486955e7319920ed69",
   "unknown",
   "Med",
   "Decontaminate first"
  ],
  [
   "S2b SFT agent/correction",
   "NuminaMath-LEAN Proof Artifacts",
   "iiis-lean/NuminaMath-LEAN-Proof-Artifacts",
   "HF-dataset",
   "corpus",
   "36,679 samples with formal statements, prover proofs and human proofs plus REPL-validated tactic and message artifacts (lite and full configs), derived from AI-MO/NuminaMath-LEAN.",
   "36679 train rows in both full and lite configs (datasets-server info); usedStorage 16895703649 bytes",
   "Lean 4.15.0 (from card, REPL runs)",
   "apache-2.0 (cardData)",
   "Verified (statement, proof) pairs and Lean REPL error messages for building correction traces. No multi-turn traces.",
   "Medium. Numina competition problems; miniF2F overlap not confirmed.",
   "API HF api 200; datasets-server info 200 2026-10-14; 43f4ea4da40c48be8e89edb84cb0825d8a4dfe2d",
   "unknown",
   "High",
   "Decontaminate first"
  ],
  [
   "S2b SFT agent/correction; S3 RL prompts",
   "MathlibPR",
   "MathlibPR/MathlibPR",
   "HF-dataset",
   "benchmark",
   "Mathlib4 pull-request histories: snapshot-level merge-ready labels (stage1-3, 15,895 snapshots each), within-PR pairs (3,687), and 500/500 subsets. It is a merge-readiness judging benchmark, not patch-writing traces.",
   "Stage configs stage1, stage2, stage3 each have 15,895 rows (test split) - candidate's per-stage-config row count confirmed. Total usedStorage 248,421,511 bytes ",
   "varies by Mathlib commit (not a single version)",
   "apache-2.0 (cardData matches candidate)",
   "Source of real Mathlib4 PR metadata (repo, PR number, commit SHAs, diffs) to build PR-style repository tasks. Needs our own task construction and a Lean build to verify.",
   "Low for miniF2F/PutnamBench statements (Mathlib source); Mathlib may include some competition-style lemmas.",
   "API 200 (HF API), 200 (datasets-server /info) 2026-10-14; 9f85f2dfe192030f58f3977fe7224b2024aed546",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S3 RL prompts",
   "DeepTheorem",
   "Jiahao004/DeepTheorem",
   "HF-dataset",
   "rl-prompts",
   "120,754 informal theorems with a truth_value flag (first-rows sample of 25 was all True), a proof, and variants",
   "120,754 rows in split train (datasets-server); download_size 554,423,240 B; dataset_size 1,185,426,199 B; usedStorage 1,702,216,290 B; 3 parquet shards plus 3 a",
   "n/a (informal, no Lean)",
   "mit (cardData and license:mit tag)",
   "Informal only, so it is not directly usable as Lean prompts. It would need autoformalizing first. The truth_value / variant design is a useful template for building prove-or-disprove pairs. Low priority.",
   "Built from informal sources such as NuminaMath. Overlap with miniF2F is unverified and likely unchecked.",
   "API 200 (HF api/datasets; datasets-server info 200) 2026-10-14; f5935720f176cedff4ecd8ebf83d1696e31cfac8",
   "https://arxiv.org/abs/2505.23754",
   "High",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "formal-conjectures (Lean 4 conjecture library)",
   "https://github.com/google-deepmind/formal-conjectures",
   "GitHub",
   "corpus",
   "Google DeepMind's Lean 4 library of formalized open and solved conjectures, many marked with answer(sorry) or true/false/undecided status, in a real repository with a Lake build",
   "unknown (statement count not checked; no tree or file queries were made)",
   "unknown (tracks recent Mathlib, likely far newer than 4.9.0-rc1)",
   "Apache-2.0 (GitHub license.spdx_id)",
   "Natural source for prove-or-disprove tasks, since the library formalizes statements whose truth value is open or recorded. It is also a real repo for PR-style environment tasks. Known-true or known-false items are the useful ones.",
   "Open-problem statements, so miniF2F, PutnamBench and ProofNet overlap should be low. Needs a modern Lean version, so incompatible with our 4.9 header without porting.",
   "API 200 (api.github.com/repos/google-deepmind/formal-conjectures) 2026-10-14; unverified (GitHub API repo endpoint does not return a commi",
   "unverified",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "Lean Workbook RL pass-rate variants (Slim205)",
   "Slim205/lean_workbook_RL_V20_total",
   "HF-dataset",
   "rl-prompts",
   "24,000 train and 434 test rows with problem_id, theorem, proof, is_proved, eval_complexity. Dozens of sibling variants (hard, interval, no_zero_examples, hinter) exist.",
   "24,434 rows",
   "unknown (Lean Workbook is 4.8.0-rc1 per other discoverer)",
   "unknown (no cardData license)",
   "Pre-bucketed Lean Workbook statements with solve flags; a template for our own pass-rate filter.",
   "Lean Workbook derives from competition-style problems; miniF2F overlap not measured.",
   "API 200 2026-10-14; sha ba8cf328; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "Lean-RLVR-dataset",
   "rootacess/Lean-RLVR-dataset",
   "HF-dataset",
   "rl-prompts",
   "100,000 rows with problem_id, formal_statement, informal_statement.",
   "100,000 rows",
   "unknown",
   "unknown (no cardData license)",
   "Statement pool for RL; provenance unknown so run decontamination.",
   "Unknown; treat as possibly overlapping with Lean Workbook and miniF2F.",
   "API 200 2026-10-14; sha da0c6e41; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "Lean4 OEIS",
   "formalica/lean4oeis",
   "HF-dataset",
   "rl-prompts",
   "OEIS-derived Lean 4 statements (100K+ files, size category 100K-1M). Schema not retrieved.",
   "100K to 1M files (size category only)",
   "unknown",
   "cc-by-sa-4.0",
   "Large synthetic statement pool; many are computational, so filter by solve rate before RL.",
   "Low for competition sets (OEIS-derived).",
   "API 200 2026-10-14; sha 279e12db; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "Lean_conjecturer_data_v02",
   "Slim205/Lean_conjecturer_data_v02",
   "HF-dataset",
   "rl-prompts",
   "95,430 train and 1,278 test rows of LLM-conjectured Lean statements (problem_id, formal_statement, proofs)",
   "95430 train + 1278 test rows (datasets-server info); usedStorage 20263355 bytes; download_size 20263355 bytes",
   "unknown",
   "unverified (no cardData license, no license tag)",
   "Synthetic conjectures with proofs attached. Provenance is uncertain, so use only after our own verification and dedup.",
   "Seeding sources are unknown. Treat as potentially overlapping with Lean Workbook and miniF2F until checked.",
   "API HF api 200; datasets-server info 200 2026-10-14; 105c06a3e6c6765b2bc1ed43046fd58a204d648e",
   "unverified",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "Mathlib refactor history (gated)",
   "princhernwang/mathlib-refactor-history",
   "HF-dataset",
   "other",
   "Mathlib refactor history, 10K-100K rows by size tag. Contents not inspected because access is manual-gated.",
   "10K to 100K rows (size category only)",
   "unknown",
   "apache-2.0",
   "Closest open analogue to repo-history PR tasks. Request access and inspect schema before planning around it.",
   "Low for miniF2F.",
   "API 200 2026-10-14; sha d18883a9; gated=manual",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "NuminaMath-LEAN-Sol and Proof-Artifacts (iiis-lean)",
   "iiis-lean/NuminaMath-LEAN-Sol",
   "HF-dataset",
   "rl-prompts",
   "81,311 rows of NuminaMath-LEAN statements with formal proofs, an English flag and match metadata. The sibling iiis-lean/NuminaMath-LEAN-Proof-Artifacts has 36,679 rows with human proof validation status and messages (full and lite configs).",
   "81311 train rows (datasets-server info); usedStorage 582466698 bytes; download_size 240385539 bytes",
   "unknown",
   "apache-2.0 (cardData)",
   "Cleaned and solved subsets of NuminaMath-LEAN. Statements with a validated proof are known provable, which removes false-statement noise from the RL pool. Proof-Artifacts' validation status helps identify items that no longer compile.",
   "Derivative of NuminaMath-LEAN, so the same miniF2F overlap risk applies.",
   "API HF api 200; datasets-server info 200 2026-10-14; d2052841575cc33d8b20bbc48b61fc8ee6a5fc0c",
   "unverified",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "RL-Goedel-Pset Level-2/3/4-5 (difficulty-bucketed)",
   "HangHor/RL-Goedel-Pset-Level-3-5",
   "HF-dataset",
   "rl-prompts",
   "Goedel-Pset subsets bucketed by an apparent difficulty level (problem_id, formal_statement, informal_statement, source such as orca_math). Sibling sets: Level-2-5 (80,382 rows), Level-4-5 (46,290 rows), both confirmed by API.",
   "63,658 rows in train split (datasets-server and hub cardData agree); train split 52,141,185 bytes per datasets-server (52,075,227 per hub cardData); download_si",
   "unknown",
   "mit (cardData.license 'mit'; tag license:mit)",
   "Pre-bucketed shortcut for the band filter on Goedel-Pset. Start with Level-3-5 and Level-4-5, then re-measure pass rate with our own SFT model.",
   "Derivative of Goedel-Pset-v1 (same risk). Provenance of the level definition is not documented in the card, so treat it as unverified.",
   "API 200 (hub API and datasets-server info both returned 200) 2026-10-14; sha c2f60f158d68d0891e653cfc87e9e618abca0eb6",
   "unverified",
   "Med",
   "Candidate"
  ],
  [
   "S3 RL prompts",
   "VeriSoftBench",
   "maxRyeery/VeriSoftBench",
   "HF-dataset",
   "benchmark",
   "500 theorem-proving tasks from 23 real Lean 4 repositories (compiler verification, type systems, ZK proofs, smart contracts), with repository context, dependency info, and ground-truth proofs",
   "500 rows in split train (datasets-server); download_size 9,215,308 B; dataset_size 8,081,374 B; verisoftbench.jsonl (usedStorage 6,842,249 B)",
   "unknown (varies per repository)",
   "mit (cardData and license:mit tag)",
   "Repo-level proving tasks with realistic cross-file dependencies. It is the nearest thing to a repository RL environment, but it is small. The GitHub repo (utopia-group/VeriSoftBench, MIT, pushed 2026-10-13) holds the pipeline and the repos. Use it as a probe set for repo-context capability.",
   "Software-verification repositories, so low miniF2F overlap. A test benchmark, so avoid training on it if it is used for eval.",
   "API 200 (HF api/datasets; datasets-server info 200) 2026-10-14; 741c2a14e68a5293b70ef874ce8b7876ce053173",
   "https://arxiv.org/abs/2602.18307",
   "Med",
   "Eval only"
  ],
  [
   "S3 RL prompts; S6 Eval (held-out)",
   "CombiBench",
   "AI-MO/CombiBench",
   "HF-dataset",
   "benchmark",
   "100 combinatorics problems formalized in Lean 4 (theorem_name, natural_language, answer, source, tag, formal_statement), with a test_with_solution split",
   "100 rows test + 100 rows test_with_solution (datasets-server info); usedStorage 608889 bytes; download_size 110669 bytes",
   "unknown (Kimina-era Lean 4)",
   "mit (cardData)",
   "Too small for RL. Use as an out-of-distribution probe for combinatorics, which is weak in Goedel-style provers.",
   "Drawn from IMO and national olympiad problems, so it may overlap with miniF2F-test IMO items. Not checked.",
   "API HF api 200; datasets-server info 200 2026-10-14; 882ba08befd0856f5364db1e53d58c7e2cf704f9",
   "https://arxiv.org/abs/2505.03171",
   "Med",
   "Eval only"
  ],
  [
   "S3 RL prompts; S6 Eval (held-out)",
   "ConjectureBench",
   "Formal-Math-Reasoning/ConjectureBench",
   "HF-dataset",
   "benchmark",
   "457 items with source, full_name, header, formal_conjecture, informal_statement, formal_statement. Conjecture-finding (answer-guessing) formalizations.",
   "457 rows in split train (datasets-server); download_size 175,147 B; dataset_size 343,889 B; usedStorage 534,323 B",
   "unknown (header column present)",
   "cc-by-4.0 (cardData and license:cc-by-4.0 tag)",
   "Small set of statements whose answer must be conjectured, a loose analogue of prove-or-disprove. Useful as a probe, not as a pool.",
   "Likely derived from existing benchmarks such as PutnamBench or competition sets. Not confirmed, so assume overlap until a check is run.",
   "API 200 (HF api/datasets; datasets-server info 200) 2026-10-14; 20d9b2b58e899889686329d5d1066e1078fa6ac4",
   "unverified",
   "Med",
   "Eval only"
  ],
  [
   "S3 RL prompts; S6 Eval (held-out)",
   "DeepSeek-ProverBench",
   "deepseek-ai/DeepSeek-ProverBench",
   "HF-dataset",
   "benchmark",
   "325 Lean 4 formalized problems (name, area, header, formal_statement), including AIME 24/25 and textbook problems",
   "325 rows train (datasets-server info); usedStorage 82194 bytes; download_size 238887 bytes",
   "unknown (a header column is present per row)",
   "unverified (cardData has no license field; no license tag)",
   "Held-out probe set, not a training pool. Useful for checking that RL on the band-filtered pool has not overfit.",
   "AIME problems overlap with miniF2F AIME items to some degree. A benchmark, so keep it out of training.",
   "API HF api 200; datasets-server info 200 2026-10-14; 3b9f067088e5e005fab91434ddc05a903e0a6252",
   "https://arxiv.org/abs/2504.21801",
   "Med",
   "Eval only"
  ],
  [
   "S3 RL prompts; S6 Eval (held-out)",
   "FormalMATH-All",
   "SphereLab/FormalMATH-All",
   "HF-dataset",
   "rl-prompts",
   "5,560 human-verified Lean 4 formalizations of olympiad-to-undergrad problems with domain labels, plus compiler feedback fields and solution",
   "5,560 rows in split train (datasets-server); download_size 14,818,432 B; dataset_size 13,356,268 B; repo files: FormalMATH_all.json, README.md. FormalMATH-Lite ",
   "unknown (paper uses Lean 4 / Mathlib; verify)",
   "mit (cardData and license:mit tag)",
   "Small, clean, fidelity-checked statements. Better as a quality dev or probe set or a high-quality RL mix-in than as bulk. The FormalMATH-Lite config (425 rows) is the ready-made eval-sized subset.",
   "Collected from competition and textbook sources, so overlap with miniF2F (AMC/AIME/IMO) is plausible. FormalMATH is itself an evaluation benchmark, so using it for training would burn it as an eval set.",
   "API 200 (HF api/datasets; datasets-server info 200) 2026-10-14; 8eea2550f174ce8b95444334da18a7473c2d169c",
   "https://arxiv.org/abs/2505.02735",
   "Med",
   "RL or eval, not both"
  ],
  [
   "S3 RL prompts; S6 Eval (held-out)",
   "miniCTX",
   "l3lab/miniCTX",
   "HF-dataset",
   "benchmark",
   "Context-dependent theorem proving from real Lean projects (PFR, HEP, HTPI, SciLean, PNT, etc.), with srcContext, theoremStatement, file, positionMetadata, dependencyMetadata, and proofMetadata. Splits are valid and test per config.",
   "6 configs, valid+test each. Split counts from datasets-server: hep 61+61, htpi 45+45, pfr 51+51, pfr_cross 43+43, pnt 85+85, scilean 46+46. Derived total (sum o",
   "unknown (varies per project; fileCreated metadata present)",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Real-repository theorems with file context. Use for repo-context RL prompts or a held-out probe. The valid splits could serve as a small RL pool.",
   "Recent project files, so low miniF2F overlap. A benchmark, so training on it removes it as an eval.",
   "API 200 (HF api/datasets; datasets-server info 200) 2026-10-14; ba24e70d112679a004510b487ebdeee8c6606ec4",
   "https://arxiv.org/abs/2408.03350",
   "Med",
   "Eval only"
  ],
  [
   "S4 Verifier/infra",
   "LeanInteract",
   "https://github.com/augustepoiroux/LeanInteract",
   "GitHub",
   "tool",
   "Python interface to the Lean REPL with multi-version Lean support. 132 stars, pushed 2026-07-17.",
   "4791 KB (repo size per GitHub API)",
   "multi-version",
   "MIT (SPDX MIT, 'MIT License')",
   "Runs compiler-feedback loops and axiom checks for trace generation and rewards.",
   "n/a",
   "API 200 2026-10-14; default_branch=main; commit sha not fetched",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "leanprover-community/repl",
   "https://github.com/leanprover-community/repl",
   "GitHub",
   "tool",
   "Official Lean 4 REPL: JSON commands over stdin/stdout, with tactic mode and environment pickling. It is the substrate under LeanInteract and Kimina-style servers.",
   "358 KB (GitHub repo size field)",
   "tracks Lean releases; tags per version (4.9.0 tag unverified)",
   "Apache-2.0 (GitHub spdx_id Apache-2.0)",
   "Base layer for the multi-turn prove-or-disprove environment and for fast proof-checking in RL. NOT a safe verifier by itself: SafeVerify's README says direct REPL checking is exploitable. Always add a SafeVerify or Comparator pass.",
   "None (tool)",
   "API GitHub /repos/leanprover-community/repl: 200 2026-10-14; unverified (default branch master; sha not fetched)",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "leanprover/Comparator",
   "https://reservoir.lean-lang.org/@leanprover/Comparator",
   "Other",
   "tool",
   "Official 'trustworthy judge for Lean proofs'. Builds Solution.lean in a landrun sandbox, exports the proof term with lean4export, and replays it outside the sandbox with the Lean kernel and optionally nanoda. Checks that the solution proves the same statement as Challenge.lean. Uses a JSON config with permitted_axioms.",
   "unknown",
   "tracks current Lean; lean4export must match the project toolchain; 4.9.0 support unverified",
   "unverified (Reservoir page, GitHub API not queried)",
   "Strongest option for the final miniF2F-test certification, and for an RL reward if throughput allows. Heavier than SafeVerify: needs landrun, lean4export, systemd-run and a non-root user. Can be used as a second, independent check in spot audits of SafeVerify-accepted proofs.",
   "None (tool)",
   "not checked",
   "https://lean-lang.org/doc/reference/4.33.0-rc1/ValidatingProofs/",
   "Low",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "leanprover/lean4checker",
   "https://github.com/leanprover/lean4checker",
   "GitHub",
   "tool",
   "Standalone replay of every declaration in an .olean through the kernel. --fresh also replays imports. This is the base that SafeVerify's Environment.replay is adapted from.",
   "160 KB (GitHub repo size field)",
   "branch per Lean release (per-version tags unverified)",
   "Apache-2.0 (GitHub spdx_id Apache-2.0)",
   "Cheap second pass. Run on accepted proofs to catch environment-manipulation exploits (e.g. addDecl tricks). Archived upstream, with the function moving into Lean core, so check the Lean version before depending on it.",
   "None (tool)",
   "API GitHub /repos/leanprover/lean4checker: 200 2026-10-14; unverified (default branch master; sha not fetched)",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "leanprover/lean4export",
   "https://reservoir.lean-lang.org/@leanprover/lean4export",
   "Other",
   "tool",
   "Exports Lean .olean environments to a text format that external kernels (nanoda and others) can re-check. It is the export step Comparator depends on.",
   "unknown",
   "must match project toolchain; per-version support unverified",
   "unverified",
   "Only needed if we adopt Comparator or independent-kernel re-checking. Must match the project's Lean version.",
   "None (tool)",
   "not checked",
   "unknown",
   "Low",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "mistralai/SafeVerify (Leanstral fork of SafeVerify)",
   "https://reservoir.lean-lang.org/@mistralai/SafeVerify",
   "Other",
   "tool",
   "Mistral's fork of SafeVerify, the verifier named in Leanstral 1.5 release materials. Listed as a Lean Reservoir package (HTTP 200 on _payload.json). The README on that page is the upstream SafeVerify README. GitHub URL not API-confirmed (github.com/mistralai/LeanstralSafeVerify was a guess and was not confirmed).",
   "unknown",
   "unknown",
   "unverified (upstream is Apache-2.0)",
   "Reference for the 'modified SafeVerify' reward gate. Read the fork's diff against upstream before copying the design. Reservoir page confirms a mistralai/SafeVerify package exists, but not what the fork changed.",
   "None (tool, no data)",
   "not checked",
   "https://mistral.ai/news/leanstral-1-5/",
   "Low",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "project-numina/kimina-lean-server",
   "https://github.com/project-numina/kimina-lean-server",
   "GitHub",
   "tool",
   "Scalable Lean verification server (REPL pool, FastAPI, cached imports). Built for RL proof-reward pipelines (Kimina-Prover).",
   "1416 KB (GitHub repo size field)",
   "unverified (configurable; docs list several)",
   "MIT (GitHub spdx_id MIT)",
   "Most direct drop-in for a high-throughput RL reward service on 4 GPUs' worth of CPU workers. Add an axiom check; the server itself only checks that proofs compile.",
   "None (tool)",
   "API GitHub /repos/project-numina/kimina-lean-server: 200 2026-10-14; unverified (default branch main; sha not fetched)",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S4 Verifier/infra",
   "HaimingW/PutnamBench-lean4",
   "HaimingW/PutnamBench-lean4",
   "HF-dataset",
   "benchmark",
   "PutnamBench Lean 4 statements. Can be used to test whether a verifier rejects sorry-filled stubs.",
   "644 rows total (train 372 + test 272; datasets-server); parquet 786,533 bytes; usedStorage 3,149,975 bytes; 4 files",
   "unknown",
   "unknown (no license in cardData; datasets-server license fie",
   "Do not use for training or RL prompts. Optional use: negative-control fixtures for the verifier.",
   "IS PutnamBench.",
   "API 200 2026-10-14; b50aeaa4e64c606b594edc3079586f11f19d2886",
   "unknown",
   "Med",
   "Eval only"
  ],
  [
   "S4 Verifier/infra",
   "leanprover/lean-eval",
   "https://reservoir.lean-lang.org/@leanprover/lean-eval",
   "Other",
   "benchmark",
   "Comparator-based formal-math benchmark. Trusted statements live in shared Lean modules, with one generated comparator workspace per problem. A problem counts as solved only if Comparator accepts. Per the search summary, lake test pins landrun, lean4export, comparator and nanoda to commits.",
   "unknown",
   "recent (tracks Tau Ceti and Mathlib pin; exact version unverified)",
   "unverified",
   "Working end-to-end template for wiring Comparator into an eval harness. Possible held-out set for RL-stage eval. Needs a recent Lean and Mathlib, so it is not usable on our 4.9.0-rc1 toolchain.",
   "Unknown. Problem set is new, but overlap with miniF2F/Putnam/ProofNet statements is unverified.",
   "not checked",
   "https://lean-lang.org/eval/",
   "Low",
   "Eval only"
  ],
  [
   "S4 Verifier/infra",
   "Pradheep1647/lean-verifier-formalizations",
   "Pradheep1647/lean-verifier-formalizations",
   "HF-dataset",
   "benchmark",
   "Lean 4 tasks for agentic coding harnesses, with fields task_statement, formal_statement, header, reference_proof, lean_toolchain, mathlib_rev, permitted_axioms ([propext, Quot.sound, Classical.choice]), repo_commit_sha, verifier_version. The first row's compile backend is 'prime-sandbox (lake env lean)'.",
   "77 rows (datasets-server size endpoint; train split only, 17 columns); parquet 3,935,438 bytes (~3.9 MB); in-memory 10,831,458 bytes (~10.8 MB); usedStorage 146",
   "v4.31.0 (first row; others unverified)",
   "multiple: mit, apache-2.0, cc-by-4.0 (cardData license list;",
   "Small test fixture for our verifier pipeline (permitted_axioms per row is the same contract as ours). Too small and too new for training. Needs Lean 4.31, so it does not run on our Lean 4.9.0 toolchain.",
   "Low for miniF2F/Putnam/ProofNet (drawn from recent arXiv-style repos), but not verified.",
   "API 200 2026-10-14; 2d581ee56e1c1b0be2d1887d39ba504e1d5e6428",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S4 Verifier/infra; S5 Bug finding",
   "mistralai/Leanstral-1.5-119B-A6B",
   "mistralai/Leanstral-1.5-119B-A6B",
   "HF-model",
   "other",
   "Model card describing the target scaffold: Mistral Vibe CLI, a lean agent config (~/.vibe/agents/lean.toml), /leanstall, and the lean-lsp-mcp. It describes serving through vLLM.",
   "usedStorage 120,943,486,394 bytes (~120.9 GB) over 13 files, including 7 consolidated safetensors shards. Per-file sizes not returned by the default API call.",
   "unknown (Lean 4 and Mathlib current)",
   "apache-2.0 (cardData.license and license tag agree)",
   "Source of truth for the agent scaffold and tool-set our traces should mimic (lean-lsp-mcp tools plus vibe). The weights are not needed for S4.",
   "N/A (model)",
   "API 200 (HF /api/models/mistralai/Leanstral-1.5-119B-A6B) 2026-10-14; 81592da95d94ab0439bfce16df1d55b402e598b6 (sha)",
   "https://mistral.ai/news/leanstral-1-5/",
   "Med",
   "Candidate"
  ],
  [
   "S4 Verifier/infra; S5 Bug finding",
   "SafeVerify",
   "https://github.com/GasStationManager/SafeVerify",
   "GitHub",
   "tool",
   "Checks a submitted proof against the reference statement and rejects non-standard axioms and exploit patterns. 49 stars, pushed 2026-04-22.",
   "165 KB (repo size per GitHub API)",
   "unknown (pinned per Lean toolchain)",
   "Apache-2.0 (SPDX Apache-2.0)",
   "Base for our reward and trace-acceptance checker.",
   "n/a",
   "API 200 2026-10-14; default_branch=main; commit sha not fetched",
   "unknown",
   "Med",
   "Candidate"
  ],
  [
   "S4 Verifier/infra; S5 Bug finding",
   "yidannwang/Leanstral-1-5-results",
   "yidannwang/Leanstral-1-5-results",
   "HF-dataset",
   "traces",
   "Preserved Leanstral-1.5 agent trajectories on miniF2F, ProofNet, FATE-M, FATE-H and PutnamBench. Includes generated Lean files, compile logs and SafeVerify logs, per-attempt metadata and REDACTIONS.md. The README reports a Putnam note where the statement-locking verifier dropped a heartbeat option, so a candidate needed independent revalidation.",
   "unknown: 67,775 files listed in siblings; usedStorage reports 0; datasets-server /info failed (config-info failed), so no row count",
   "unknown",
   "unknown (no license in cardData; no license field in the API",
   "Reference for verifier behavior, including edge cases and failure logs. Useful for calibrating our SafeVerify wrapper against real accept and reject cases. NOT for training: it contains eval benchmark artifacts.",
   "HIGH. Contains miniF2F, ProofNet and PutnamBench statements, proofs and trajectories. Exclude from any training mix.",
   "API 200 2026-10-14; f2d4b411a7d298bb5ac8f5b189a79a704f1f95ff",
   "https://mistral.ai/news/leanstral-1-5/",
   "Med",
   "Decontaminate first"
  ],
  [
   "S5 Bug finding",
   "Aeneas",
   "https://github.com/AeneasVerif/aeneas",
   "GitHub",
   "tool",
   "Rust-to-pure-functional translation backend; emits Lean 4 (also Coq/F*/HOL4) from Charon's LLBC. Core tool of the Leanstral bug-finding pipeline.",
   "unknown (tool; repo size 19,163 KB as reported by GitHub API)",
   "unverified (tracks a pinned Lean 4 toolchain; read lean-toolchain in the repo; likely much newer than Goedel's 4.9.0-rc1)",
   "Apache-2.0 (license.spdx_id)",
   "Run on Rust repos to get Lean models for the property and negation proving loop. This is the only way to build our own bug-finding set.",
   "None (tool)",
   "API 200 2026-10-14; unverified (repos API response does not expose a commit sha;",
   "https://arxiv.org/abs/2206.07185",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "Charon",
   "https://github.com/AeneasVerif/charon",
   "GitHub",
   "tool",
   "Rust MIR front-end that extracts ULLBC/LLBC from cargo crates. It feeds Aeneas and other backends.",
   "unknown (tool; repo size 20,078 KB as reported by GitHub API)",
   "n/a (front-end only)",
   "Apache-2.0 (license.spdx_id)",
   "Step 1 of Rust to Lean: extract LLBC from each target repo.",
   "None (tool)",
   "API 200 2026-10-14; unverified (no commit sha in repos API response; default bra",
   "https://arxiv.org/abs/2410.18042",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "dafny-disco",
   "metareflection/dafny-disco",
   "HF-dataset",
   "corpus",
   "81,342 Dafny content items grouped by id/group (about 518 MB). Looks like a large Dafny corpus (I did not inspect the rows).",
   "81,342 rows, split train (datasets-server); dataset_size 504,789,705 B (~504.8 MB); download 114,913,745 B (~114.9 MB); 2 parquet shards",
   "n/a (Dafny)",
   "cc-by-sa-4.0 (cardData and license tag agree)",
   "A possible source of verified-program text for the translation direction. Low priority because it is not Lean.",
   "None for Lean eval sets.",
   "API 200 (api/datasets); datasets-server /info OK 2026-10-14; 4fc7b5adfe36c8402a718311be3e1b3d225530f1",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "DafnyGym",
   "emugnier/DafnyGym",
   "HF-dataset",
   "rl-prompts",
   "145 Dafny methods with assertions removed, plus .dfy files in assertions_removed/. An environment for repairing proofs in Dafny.",
   "145 rows, split train, config main_data (datasets-server); dataset_size 328,873 B, download 368,334 B; usedStorage 80,832 B; 539 sibling files (.dfy per-lemma f",
   "n/a (Dafny)",
   "mit (cardData and license tag agree)",
   "A reference for building a verifier-in-the-loop 'gym' for program-proof tasks. It is not directly usable for Lean.",
   "None for Lean eval sets.",
   "API 200 (api/datasets); datasets-server /info OK 2026-10-14; 44aa7efe3fc8e5228405d6337270d6968be75b40",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "FVAPPS (Formally Verified APPS)",
   "quinn-dougherty/fvapps",
   "HF-dataset",
   "sft-data",
   "Lean 4 theorem statements with sorry for property-style specs over APPS programming problems. Files are train.jsonl, test.jsonl, assured.jsonl and semiassured.jsonl.",
   "exact row count unverified (datasets-server 401). Files: assured.jsonl, semiassured.jsonl, test.jsonl, train.jsonl, unassured.jsonl; usedStorage 95,844,962 byte",
   "unverified (Lean 4; the repo may be pinned to an older Mathlib)",
   "mit (cardData.license)",
   "Source of Lean property statements, with the program and tests as context, for prove-property training or RL prompts. The Lean programs are real code properties, not math.",
   "Low (APPS-derived). No miniF2F, PutnamBench or ProofNet overlap found.",
   "API 200 (api/datasets); datasets-server /info 401 (gated/auth required) 2026-10-14; 3552a173e74837dd0017d72431c6ede8fa903bb7",
   "https://arxiv.org/abs/2502.05714",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "hax",
   "https://github.com/cryspen/hax",
   "GitHub",
   "tool",
   "Rust verification toolchain. It translates a Rust subset to F*, Rocq and Lean backends.",
   "unknown (tool; repo size 74,025 KB as reported by GitHub API)",
   "unverified",
   "Apache-2.0 (license.spdx_id)",
   "Alternative or complement to Aeneas for Rust to Lean. Its Lean backend maturity is unverified.",
   "None (tool)",
   "API 200 2026-10-14; unverified (no commit sha in repos API response; default bra",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "RLVR code data (Lean translation)",
   "saurabh5/rlvr-code-data-Lean",
   "HF-dataset",
   "rl-prompts",
   "133,285 rows of translated programming problems with solutions and test cases, target_language Lean.",
   "133,285 rows",
   "unknown",
   "unknown (no cardData license)",
   "Program-writing tasks in Lean with tests; usable for verified-code RL, not for theorem proving.",
   "Low for math benchmarks.",
   "API 200 2026-10-14; sha cd40efdf; gated=False",
   "none",
   "Med",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "RustSec advisory-db",
   "https://github.com/rustsec/advisory-db",
   "GitHub",
   "corpus",
   "Machine-readable advisories for vulnerabilities and unsound code in crates.io crates, in TOML with affected versions, CVE/GHSA aliases and patched versions.",
   "repo size 41851 kB (GitHub API 'size' field); advisory count not counted (unknown)",
   "n/a",
   "NOASSERTION (GitHub API license.spdx_id; license key 'other'",
   "Ground truth for a held-out bug-finding eval. Pick crates with known soundness or logic bugs, run Charon and Aeneas on the pre-fix version, and check whether property or negation proving finds the bug.",
   "None for Lean evals. Possible overlap with public bug lists the base model has seen.",
   "API 200 2026-10-14; unknown (default branch main; sha not fetched)",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "Verus_Training_Data (SAFE / VeruSyn)",
   "microsoft/Verus_Training_Data",
   "HF-dataset",
   "sft-data",
   "Training data for Verus (verified Rust). SAFE SFT 25K samples (11K proof generation, 14K debugging); a 6.9M-sample SFT part1; 4.6K long-CoT part2; and raw VeruSyn trajectories (success 7396, error 271, timeout 1373).",
   "Per-file bytes (blobs=true): sft_part1_6.9M.json 15,681,162,022 B; sft_part2_4557.json 95,055,075 B; sft_safe_25k.json 59,123,561 B; system_trajectory_843.jsonl",
   "n/a (Verus)",
   "unknown (cardData license null; no license tag) - confirmed",
   "The only large verified-Rust corpus. Use it for debugging and repair trajectory shapes, or to bootstrap understanding of Rust code and specs. Not Lean, so it is a design reference.",
   "None for Lean math evals.",
   "API 200 (api/datasets); datasets-server /info FAILED (config-info failed, dataset_info empty) 2026-10-14; d3a283fabb152ceee6e3e350a30d37499ece4f90",
   "https://arxiv.org/abs/2410.15756",
   "Med",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "VeruSyn",
   "anon-verusyn/verusyn",
   "HF-dataset",
   "sft-data",
   "A mirror of VeruSyn SFT parts: part1 (6.9M) and part2 (4,557 samples with prompt/input/output).",
   "part2: 4,557 rows, split train (datasets-server, config part2), dataset_size 91,450,596 B; files: sft_part1_6.9M.json 15,681,162,022 B (~15.7 GB, row count UNVE",
   "n/a (Verus)",
   "mit (cardData and license tag agree; datasets-server info li",
   "An alternative handle on the same verified-Rust data. Same use as Verus_Training_Data.",
   "None for Lean math evals.",
   "API 200 (api/datasets); datasets-server /info OK for config part2 only 2026-10-14; 73ef2bc851358e513418df8a5ed5b551b3a417de",
   "unknown",
   "High",
   "Candidate"
  ],
  [
   "S5 Bug finding",
   "CLEVER (dataset)",
   "amitayusht/clever",
   "HF-dataset",
   "benchmark",
   "161 hand-curated HumanEval-derived tasks. Each has an NL spec, a formal ground-truth spec, an isomorphism theorem and a Lean implementation task (spec-certification plus implementation-certification).",
   "161 train examples (datasets-server); download_size 340,602 bytes; dataset_size 341,241 bytes; usedStorage 151,034 bytes",
   "unverified (see GitHub repo for toolchain)",
   "mit (cardData.license; matches card)",
   "Dev or eval set for 'write a Lean spec or property and prove it'. Also useful as a dev set for property generation, since a spec equals a property.",
   "Low for miniF2F/PutnamBench/ProofNet (code tasks derived from HumanEval). Overlaps with HumanEval-derived code benchmarks.",
   "API 200 (api/datasets); datasets-server /info 200 2026-10-14; 486ea1554eb7ce6104e35a138e60174cdeb08c7b",
   "https://arxiv.org/abs/2505.13938",
   "High",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "CodeWalk Verus proof completion",
   "CodeWalk/verus-proof-completion",
   "HF-dataset",
   "benchmark",
   "400 Verus (Rust) proof-completion problems at 4 chain levels (L1, L2, L3, L10, 100 each). Inputs only, no solutions.",
   "unverified rows (datasets-server reported no splits); HF siblings: 411 files, sizes not returned by API; card claim of 400 problems not confirmed by API",
   "n/a (Verus)",
   "mit (cardData.license and license:mit tag)",
   "Reference design for a multi-function proof benchmark. Not Lean.",
   "None for Lean evals. The card carries a canary GUID: keep it out of training corpora.",
   "API api 200; datasets-server info 200 but dataset_info empty (no splits reported) 2026-10-14; f0e6ec27f39a688ac130d2e98b386ea82f3098b4",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "Dafny-synthesis (MBPP-DFY)",
   "https://github.com/Mondego/dafny-synthesis",
   "GitHub",
   "benchmark",
   "MBPP-DFY-153: Dafny programs with specs and proofs generated from MBPP. Cited from memory as a starting hint.",
   "unverified (about 153 per candidate name; not confirmed)",
   "n/a (Dafny)",
   "unverified (API not retrieved)",
   "A Dafny-side reference for spec and proof generation on small programs.",
   "None for Lean evals.",
   "Missing (404)",
   "https://arxiv.org/abs/2402.00247",
   "Low",
   "Missing (404)"
  ],
  [
   "S5 Bug finding",
   "DafnyBench",
   "wendy-sun/DafnyBench",
   "HF-dataset",
   "benchmark",
   "782 Dafny programs with verification hints (assertions, invariants) removed. The model must restore them so Dafny verifies.",
   "782 rows, split test (datasets-server); files: test.json; dataset_size 2,816,027 B, download 3,014,671 B; usedStorage 9,945,963 B",
   "n/a (Dafny)",
   "apache-2.0 (cardData and license tag agree)",
   "Not Lean. Use only as a program-verification task-shape reference or to transfer to Lean by translating. It is an indirect proxy for 'prove the property'.",
   "None for the Lean eval sets.",
   "API 200 (api/datasets); datasets-server /info OK 2026-10-14; dec1a9774b95689dd5b7b741cdc0366e966433de",
   "https://arxiv.org/abs/2406.08467",
   "High",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "DafnyComp",
   "ReFormDafny/DafnyComp",
   "HF-dataset",
   "benchmark",
   "300 compositional Dafny verification tasks (fields: org_input, gt_output, dafny_input, stage_tag, prompt, ability).",
   "300 rows, split test (datasets-server); dataset_size 7,863,189 B; download 2,326,883 B; single test.parquet",
   "n/a (Dafny)",
   "unknown (cardData has no license; no license tag)",
   "A reference for compositional (multi-function) spec checking, similar to repo-level property proving. Not Lean.",
   "None for Lean eval sets.",
   "API 200 (api/datasets); datasets-server /info OK 2026-10-14; c574d71b2f022d7a62b6642ae9a4e4b799cbeb4f",
   "unknown",
   "Med",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "miniCodeProps",
   "elohn/miniCodeProps",
   "HF-dataset",
   "benchmark",
   "A Lean 4 benchmark of 201 properties of simple programs (sorts, lists, trees, etc.) with full-proof and next-step baselines, mined from Tons-of-Theorems-style code.",
   "n<1K (size_categories tag confirmed); row count unverified (single file codeprops_bench_ps.jsonl, usedStorage 41,937 bytes); 201 properties not confirmed via AP",
   "unverified (the repo carries an older Lean and Mathlib; likely 4.x mid-2024)",
   "bsd-3-clause (cardData.license)",
   "Dev or eval for proving program properties and their negations. It is small and CPU-cheap to check.",
   "Low (program properties, not competition math).",
   "API 200 (api/datasets); datasets-server /info 200 but failed config-info (no size) 2026-10-14; a11d7e216e4beb5dcf580edc06624e0baa95fdb7",
   "https://arxiv.org/abs/2406.11915",
   "High",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "Vericoding benchmark (Dafny / Verus / Lean)",
   "beneficial-ai-foundation/vericoding",
   "HF-dataset",
   "benchmark",
   "12,504 formal specs across languages (search snippet: 3,029 Dafny, 2,334 Verus/Rust, 7,141 Lean). Columns include language, source, vc-description, vc-preamble, vc-helpers.",
   "12,504 rows",
   "unknown (Lean likely needs a newer toolchain than 4.9.0-rc1)",
   "mit",
   "Lean spec-to-verified-code tasks for the property-proving and bug-finding stage. Also usable as RL prompts for program-verification.",
   "Low for miniF2F/PutnamBench/ProofNet; these are program specs, not competition math.",
   "API 200 2026-10-14; sha f14842db; gated=False",
   "unverified",
   "Med",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "VerinaLite",
   "XiaoyangLiu-sjtu/VerinaLite",
   "HF-dataset",
   "benchmark",
   "A reduced variant of Verina produced by the VeriScale expansion-and-reduction pipeline. Per-task folders hold description.txt, reject_inputs.json, signature.json and task.json.",
   "unverified (folder layout verina_advanced_N/{description.txt, reject_inputs.json, signature.json, task.json, task.lean, test.json}; datasets-server has no confi",
   "unknown",
   "apache-2.0 (cardData.license)",
   "A smaller, cheaper Verina slice for fast dev loops on a single A10G.",
   "Low (a Verina derivative). The README is a one-liner.",
   "API 200 (api/datasets); datasets-server /info 200 but failed config-info (no size) 2026-10-14; 5987e0926161447893ed047191497ee4c37089e7",
   "https://arxiv.org/abs/2605.22368",
   "Med",
   "Eval only"
  ],
  [
   "S5 Bug finding",
   "VerinaPlus",
   "XiaoyangLiu-sjtu/VerinaPlus",
   "HF-dataset",
   "benchmark",
   "An expanded Verina variant from the VeriScale pipeline, with more tasks per original problem. Same per-task folder layout as VerinaLite.",
   "unverified (folder layout verina_advanced_N/...; datasets-server has no config info; task count not counted)",
   "unknown",
   "apache-2.0 (cardData.license)",
   "A larger set of Verina-style tasks. It could serve as RL prompts after filtering to 'rarely but not never solved'.",
   "Low (derived from Verina). Leakage into Verina eval is possible if the same base problems are used.",
   "API 200 (api/datasets); datasets-server /info 200 but failed config-info (no size) 2026-10-14; e21184af802c198d94cd68a6de447104fcfb00db",
   "https://arxiv.org/abs/2605.22368",
   "Med",
   "Eval only"
  ],
  [
   "S5 Bug finding; S6 Eval (held-out)",
   "Verina",
   "sunblaze-ucb/verina",
   "HF-dataset",
   "benchmark",
   "189 Lean 4 verifiable-code-generation tasks. Each has an NL description, a reference Lean implementation, a formal specification, a proof, test cases and reject inputs, in two difficulty tiers (basic/advanced).",
   "189 train examples (datasets-server); download_size 214,167 bytes; dataset_size 613,699 bytes; usedStorage 1,770,315 bytes; 1 parquet file (data/train-00000-of-",
   "unverified (check the GitHub repo sunblaze-ucb/verina)",
   "apache-2.0 (cardData.license)",
   "Closest benchmark to 'generate property + prove it'. Use it to measure proving specs and implementations in Lean, and as a source of (spec, proof) pairs for SFT or RL warm-up.",
   "Low (programming tasks, derived from MBPP and LiveCodeBench-style sources). It has no miniF2F or PutnamBench statements.",
   "API 200 (api/datasets); datasets-server /info 200 2026-10-14; da391829470768353e02cfc9ca03f492ce109b58",
   "https://arxiv.org/abs/2505.23135",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "FATE (FrenzyMath) repo",
   "https://github.com/frenzymath/FATE",
   "GitHub",
   "benchmark",
   "FATE-M/H/X abstract-algebra and commutative-algebra Lean4 formalizations of increasing difficulty",
   "unverified (GitHub repo size field 14 KB; not counted, as instructed)",
   "unverified",
   "MIT (license.spdx_id MIT)",
   "Hard research-level eval where the baseline is near zero. Treat as a low-signal sanity test.",
   "New, low overlap expected.",
   "API 200 (GitHub api) 2026-10-14; unknown (default branch main; commit sha not fetched)",
   "https://arxiv.org/abs/2505.17262 (unverified id)",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "FormalMATH-Lite",
   "SphereLab/FormalMATH-Lite",
   "HF-dataset",
   "benchmark",
   "425-problem subset of FormalMATH (Lean4, olympiad to undergrad)",
   "425 rows (train split, num_examples 425); dataset_size 833190 bytes; download_size 940415 bytes; usedStorage 343795 bytes",
   "unverified",
   "mit (cardData.license and tag)",
   "Cheap broad eval for A10G budgets. Prefer Lite over All.",
   "Sourced from olympiad and textbook problems. Overlap with miniF2F possible. Dedup before use.",
   "API 200 (HF api and datasets-server info) 2026-10-14; 75e91a83573b6241fb85fe2a9cef8dd6b713f170",
   "https://arxiv.org/abs/2505.02735",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "Lean-eval benchmark (formalmathatepfl)",
   "formalmathatepfl/lean-eval-benchmark",
   "HF-dataset",
   "benchmark",
   "Large Lean eval set, 68,932 test rows, Lean v4.15.0 per card",
   "68932 rows (test split, num_examples 68932); dataset_size 335050623 bytes; download_size 33205069 bytes; usedStorage 33205069 bytes",
   "v4.15.0 (card)",
   "none declared (no license tag and no cardData.license)",
   "Possible big held-out pool. Provenance unclear, so treat as unvetted.",
   "Unknown. Probably overlaps public corpora.",
   "API 200 (HF api and datasets-server info) 2026-10-14; 5a673224433f394901399517d9311a0c72e85ea2",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "miniCTX-v2",
   "l3lab/miniCTX-v2",
   "HF-dataset",
   "benchmark",
   "Updated miniCTX with ConNF, FLT, HepLean, Seymour, carleson, foundation, mathlib (50 valid and 50 test each, FLT 34/34)",
   "334 rows per split: valid 334 and test 334 across 7 configs (ConNF 50/50, FLT 34/34, HepLean 50/50, Seymour 50/50, carleson 50/50, foundation 50/50, mathlib 50/",
   "varies per project",
   "apache-2.0 (cardData.license and tag)",
   "Time-split repo-context eval. FLT split gives a tiny FLTEval-like probe.",
   "Repos postdate many corpora. Check cutoffs against S1 data.",
   "API 200 (HF api and datasets-server info) 2026-10-14; 91bd27f994c6fd6e3e1a85e7ad01c4ee0e6a01de",
   "https://arxiv.org/abs/2408.03350",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "miniF2F Lean4 (yangky11 repo)",
   "https://github.com/yangky11/miniF2F-lean4",
   "GitHub",
   "benchmark",
   "Lean4 miniF2F Lake project used by LeanDojo/ReProver ecosystems",
   "unverified: API gives repo size only (397 KB); 244+244 statement count not recounted",
   "unverified (see lean-toolchain in repo)",
   "MIT (GitHub license.spdx_id: MIT)",
   "Reference for header and Lean toolchain pin if statements need re-checking.",
   "IS miniF2F.",
   "API 200 2026-10-14; default branch main; commit sha not returned by repos endpoi",
   "https://arxiv.org/abs/2306.15626",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "MiniF2F traces (kobyjl)",
   "kobyjl/mini_f2f_traces",
   "HF-dataset",
   "traces",
   "3,940 attempt records on miniF2F challenges (challenge_id, theorem_name, challenge_lean, prover_id, attempt, stop_reason).",
   "3,940 rows",
   "unknown",
   "mit",
   "Reference for baseline comparison only. Add to the decontamination blocklist.",
   "DIRECT: contains miniF2F statements and attempts.",
   "API 200 2026-10-14; sha e6fb8501; gated=False",
   "none",
   "Med",
   "Exclude (eval-derived)"
  ],
  [
   "S6 Eval (held-out)",
   "miniF2F-test (AI-MO / Kimina)",
   "AI-MO/minif2f_test",
   "HF-dataset",
   "benchmark",
   "miniF2F test split 244, as used in Kimina-Prover eval",
   "244 rows (train split, datasets-server num_examples); data/train-00000-of-00001.parquet, download_size 74285 bytes, dataset_size 156154 bytes; usedStorage 17081",
   "unverified",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Cross-check statement set against cat-searcher copy; diff to catch port differences.",
   "IS miniF2F test.",
   "API api 200; datasets-server info 200 2026-10-14; 5def318348521dfc34875045a9ecddf729a2b49f",
   "https://arxiv.org/abs/2504.11354",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "miniF2F-test Lean4 (HaimingW)",
   "HaimingW/miniF2F-test-lean4",
   "HF-dataset",
   "benchmark",
   "miniF2F test Lean4 244 rows",
   "244 rows (train split, datasets-server num_examples); download_size 134138 bytes, dataset_size 280937 bytes; usedStorage 487366 bytes",
   "unverified",
   "none declared (cardData.license null; no license: tag)",
   "Third-party copy for cross-checking only. Prefer the license-clear copies.",
   "IS miniF2F test.",
   "API api 200; datasets-server info 200 2026-10-14; 7046e9d7f0de6d839013e4bb463835d273d4acc7",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "miniF2F-v2 (corrected, v2s/v2c)",
   "roozbeh-yz/miniF2F_v2",
   "HF-dataset",
   "benchmark",
   "Corrected miniF2F with simplified (v2s) and competition-style (v2c) jsonl files, fixing statement/informal mismatches",
   "unverified row counts (datasets-server returned no info); 2 jsonl files confirmed: miniF2F_v2c.jsonl and miniF2F_v2s.jsonl; per-file sizes not returned by API (",
   "unverified",
   "apache-2.0 (cardData and license:apache-2.0 tag)",
   "Secondary robustness eval, reported separately from the headline miniF2F-test number.",
   "Derivative of miniF2F. Exclude.",
   "API api 200; datasets-server info 200 with failed config-info for config default (dataset_info empty) 2026-10-14; 065cb8057ccfc6e70f4afdd4d8542306c0f7cdb2",
   "unknown (check card)",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "ProofNet# (ProofNetSharp, Lean 4)",
   "PAug/ProofNetSharp",
   "HF-dataset",
   "benchmark",
   "Corrected Lean 4 ProofNet#, valid 185 / test 186",
   "371 rows (valid 185 + test 186, datasets-server)",
   "compatible v4.7.0 to v4.16.0-rc2 (card)",
   "mit (cardData.license; top-level license field null)",
   "Preferred out-of-domain (undergrad) eval, proofnet test only for final.",
   "IS ProofNet. Exclude.",
   "API 200 2026-10-14; a8da405fbd1e348a87445c2e562c747b7e26dc8f",
   "https://arxiv.org/abs/2406.xxxxx (unverified; ProofNet# / PAug)",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "ProofNet-lean4 (UDACA)",
   "UDACA/proofnet-lean4",
   "HF-dataset",
   "benchmark",
   "ProofNet Lean 4 port, 185 valid / 186 test; v2 and v3 variants exist (UDACA/proofnet-v2-lean4, proofnet-v3-lean4)",
   "371 rows (validation 185 + test 186, datasets-server)",
   "unverified",
   "none declared (no cardData.license, no top-level license)",
   "Alternate port to compare against ProofNet#.",
   "IS ProofNet.",
   "API 200 2026-10-14; 947c7189d853a3d0c2fffd25959ff3a3548b55e2",
   "https://arxiv.org/abs/2302.12433",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "ProofNet-Verified",
   "marcusm117/ProofNet-Verified",
   "HF-dataset",
   "benchmark",
   "Verified/cleaned ProofNet statements, 367 rows",
   "367 rows (train split, datasets-server)",
   "unverified",
   "apache-2.0 (cardData.license)",
   "Optional cleaned variant. Check what was verified before adopting.",
   "Derivative of ProofNet.",
   "API 200 2026-10-14; 474192c2364ee646cbb5314c44dbecd48ae89d92",
   "unknown",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "PutnamBench (lmms-lab mirror, Lean4/Isabelle/Coq)",
   "lmms-lab/PutnamBench",
   "HF-dataset",
   "benchmark",
   "PutnamBench with lean4 config (660 rows), isabelle (640), coq (410)",
   "lean4 test 660 rows (676,720 bytes); isabelle test 640; coq test 410 (datasets-server)",
   "unverified (upstream tracks recent Lean; likely newer than 4.9)",
   "none declared (no cardData.license)",
   "Hard OOD eval at pass@k. Expect near-zero for 8B models. Use only the lean4 config.",
   "IS PutnamBench. Exclude.",
   "API 200 2026-10-14; cd981a0a118b2c1152fec5861e282388e2391b05",
   "https://arxiv.org/abs/2407.11214",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "PutnamBench (trishullab upstream)",
   "https://github.com/trishullab/PutnamBench",
   "GitHub",
   "benchmark",
   "Canonical PutnamBench repo, Lean4/Isabelle/Coq formalisations",
   "unknown: repo size 3503 KB; statement count not returned by API, so the 522 to 672 range is unverified",
   "unverified (check lean-toolchain per commit)",
   "null (GitHub license.spdx_id and name both null; no top-leve",
   "Pin a commit hash for the statement set and Lean version. Needed because mirrors differ in counts.",
   "IS PutnamBench.",
   "API 200 2026-10-14; default branch main; commit sha not returned (unknown)",
   "https://arxiv.org/abs/2407.11214",
   "High",
   "Eval only"
  ],
  [
   "S6 Eval (held-out)",
   "Seed-Prover (ByteDance)",
   "https://github.com/ByteDance-Seed/Seed-Prover",
   "GitHub",
   "benchmark",
   "Seed-Prover release with IMO-style and miniF2F/PutnamBench results and generated proofs",
   "unverified (GitHub repo size field 14034 KB is the repo size, not a dataset size)",
   "unverified",
   "Apache-2.0 (license.spdx_id Apache-2.0)",
   "Reference for IMO formal statements and reproducing reported numbers. Statement files not confirmed.",
   "Contains miniF2F/Putnam proofs. Never use in training.",
   "API 200 (GitHub api) 2026-10-14; unknown (default branch main; commit sha not fetched)",
   "https://arxiv.org/abs/2507.23726",
   "High",
   "Eval only"
  ]
 ]
}