task_id stringlengths 38 100 | task_type stringclasses 1
value | theorem_statement stringlengths 120 11.9k | target_workspace dict | difficulty stringclasses 3
values | source stringclasses 4
values | source_id stringlengths 10 74 | source_license stringclasses 1
value | successful_trajectories listlengths 1 3 |
|---|---|---|---|---|---|---|---|---|
nmp3_0001e8eb-17b3-5bd9-b7ec-fa259a6b2a02 | lean_theorem_proving | "-- Original natural-language problem:\n-- Determine all functions $f:\\mathbb{R}_{\\ge0}\\to\\mathb(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 0001e8eb-17b3-5bd9-b7ec-fa259a6b2a02 | cc-by-4.0 | [{"id":"65385af355fe49e3994e7c6aaea79594","problem_id":"7f9febaac23d26bd018bf003e51d29c4c7e3a9c52682(...TRUNCATED) |
nmp3_0038fcfc-14de-5989-bd2e-befab5715d65 | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $f:[a,b]\\to[a,b]$ be a continuous function. Suppose (...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 0038fcfc-14de-5989-bd2e-befab5715d65 | cc-by-4.0 | [{"id":"e2c209b3e5254a51bb16111b9243c391","problem_id":"6ea87a662d2ba6a84deb93f0ab03c3dbb3da07389d43(...TRUNCATED) |
nmp3_00433297-225c-5a9a-9365-f121df4aee70 | lean_theorem_proving | "-- Original natural-language problem:\n-- Suppose $a,b,c$ are real numbers with $a+b+c=3$. Prove th(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 00433297-225c-5a9a-9365-f121df4aee70 | cc-by-4.0 | [{"id":"3b36a6efac0946d4ab59115054059750","problem_id":"7e179a2c994c20ef45afe9e34a5d5d04daf36cd6dcbf(...TRUNCATED) |
nmp3_005bdb6d-a4be-59c4-90a2-b42904244471 | lean_theorem_proving | "-- Original natural-language problem:\n-- The $2001$ towns in a country are connected by some roads(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 005bdb6d-a4be-59c4-90a2-b42904244471 | cc-by-4.0 | [{"id":"cb6209dc6e95447bbd611aaff39e90d4","problem_id":"306f4bbf741abc935b3ddbc75491e176356b02af2f50(...TRUNCATED) |
nmp3_0060bc23-6a89-55a0-a2e2-51056a34aae9 | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $a,b,c\\ge 0$ satisfy $a^{2}b+b^{2}c+c^{2}a=1$. Prove(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | hard | nvidia/Nemotron-Math-Proofs-v3-RL | 0060bc23-6a89-55a0-a2e2-51056a34aae9 | cc-by-4.0 | [{"id":"d9ecd56310e844108d6b4f4536f6bb96","problem_id":"b14c5f0a430731a7adf17a803ed2668051b2ab1e34ec(...TRUNCATED) |
nmp3_006262ed-772c-5620-b3fe-20a443b17b3d | lean_theorem_proving | "-- Original natural-language problem:\n-- Every integer greater than 1000 is colored either red o(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 006262ed-772c-5620-b3fe-20a443b17b3d | cc-by-4.0 | [{"id":"7e678e5ba5d94dbba4f66de7f7ec525a","problem_id":"c1a311f0ee7d0633048b4d9b63c9cb0d058a4c34b45b(...TRUNCATED) |
nmp3_00636aa7-4b2f-52ba-8da3-8973991f2ead | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $a,b,c>0$ be real numbers such that $a^{2}+b^{2}+c^{2(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 00636aa7-4b2f-52ba-8da3-8973991f2ead | cc-by-4.0 | [{"id":"7d6fcbf40a12403d942c117d452bb7bd","problem_id":"ab93e2e5cb83972f43926931bb700ae7175812a0d128(...TRUNCATED) |
nmp3_009daa63-9878-5893-b975-ac888618b565 | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $a,b,c$ be positive real numbers such that $a+b+c=3$.(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | hard | nvidia/Nemotron-Math-Proofs-v3-RL | 009daa63-9878-5893-b975-ac888618b565 | cc-by-4.0 | [{"id":"e739bfad4c5d4e49bf723c2849093be4","problem_id":"f3115d748215a33a43a2ee5fa7fbb265e29a5e780822(...TRUNCATED) |
nmp3_00ac1c64-6c7b-5f4a-a0d1-f61461ca525d | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $n\\ge 3$ be an integer. There are $n$ people and exa(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 00ac1c64-6c7b-5f4a-a0d1-f61461ca525d | cc-by-4.0 | [{"id":"39ef891b5d634fa8af962a98ca0eb16c","problem_id":"e9ab1afd96944098d147da8fc1e4bcff0d60f938d2d1(...TRUNCATED) |
nmp3_00ac344b-9d0b-5051-87c0-b106c9048e51 | lean_theorem_proving | "-- Original natural-language problem:\n-- Let $a,b,n$ be integers with $a>b>1$, $b$ odd, and $n\\ge(...TRUNCATED) | {"name":"target","commit_hash":"56a005855ea17e0b0958f73ec801ae140b0a1679","repo_url":"https://github(...TRUNCATED) | medium | nvidia/Nemotron-Math-Proofs-v3-RL | 00ac344b-9d0b-5051-87c0-b106c9048e51 | cc-by-4.0 | [{"id":"9a4b7d78a023483e9bb1f48ef690ae1d","problem_id":"8dd3d1817756c91a202352767817c2e9c0b7a9e1b843(...TRUNCATED) |
Agentic Prover RL
4,030 Lean 4 theorem-proving tasks for agentic RL. Each task has at least one solution generated with DeepSeek V4.1 Flash and accepted by daVinci-prover under Lean 4.35.0.
Sources
- 2,266 tasks: our Lean formalizations of selected problems from NVIDIA's nvidia/Nemotron-Math-Proofs-v3-RL, whose source problems are licensed under CC BY 4.0.
- 1,764 tasks: our filtered oProof/Goedel pool. Original source identifiers are retained.
Load
from datasets import load_dataset
train = load_dataset("davinci-prover/agentic-prover-RL", split="train")
trajectory = train[0]["successful_trajectories"][0]
messages = trajectory["messages"]
Data format
| Field | Description |
|---|---|
task_id, task_type |
Unique task ID and lean_theorem_proving. |
theorem_statement |
Lean task source with the proof goal. |
target_workspace |
Pinned Mathlib repository and Lean toolchain. |
difficulty |
easy (1,313), medium (1,696), or hard (1,021). |
source, source_id, source_license |
Upstream provenance; license is null when absent from source metadata. |
successful_trajectories |
List of verified trajectories in the SFT row format: messages, tools, source_files, metadata, identifiers, and model/environment versions. |
Difficulty is a teacher-cost heuristic based on output tokens, interaction/Lean calls, and valid failures, excluding infrastructure errors. Tasks are deduplicated by ID; the split is train.
4,115 successful trajectories are attached to the 4,030 tasks; multiple successes are retained in chronological order. messages preserves the final saved conversation, including reasoning_content, tool calls, and feedback. tools, metadata, and tool-call function.arguments are JSON strings. source_files includes the accepted proof and required helper modules with SHA-256 hashes.
These are successful evaluation records; SFT filtering has not been applied. History recovery and budget-stopped successes are retained. Recovery records are in json.loads(trajectory["metadata"])["dvp_rl"]["rejected_responses"]; termination and tool-call completion flags are recorded alongside them.
Citation
@article{zhu2026davinci,
title={daVinci-Prover: Agent-Native Mid-Training for Formal Theorem Proving},
author={Zhu, Wenhong and Li, Xuefeng and Zhang, Jiashuo and Chen, Yu and Mi, Tiantian and Guo, Qipeng and Liu, Pengfei},
journal={arXiv preprint arXiv:2611.16433},
year={2026}
}
- Downloads last month
- 156