wh-zhu commited on
Commit
b91e231
·
verified ·
1 Parent(s): 9f2cec4

Add 4,030 verified Lean RL tasks with difficulty labels

Browse files

Publish the fixed 4,030-task snapshot with Lean 4.35.0 workspace metadata and source attribution, including 2,266 formalized NVIDIA Nemotron-Math-Proofs-v3-RL problems.

Files changed (2) hide show
  1. README.md +52 -0
  2. data/train.parquet +3 -0
README.md ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Agentic Prover RL
3
+ language:
4
+ - en
5
+ tags:
6
+ - lean4
7
+ - theorem-proving
8
+ - mathematics
9
+ - agentic
10
+ - reinforcement-learning
11
+ size_categories:
12
+ - 1K<n<10K
13
+ task_categories:
14
+ - text-generation
15
+ source_datasets:
16
+ - nvidia/Nemotron-Math-Proofs-v3-RL
17
+ configs:
18
+ - config_name: default
19
+ default: true
20
+ data_files:
21
+ - split: train
22
+ path: data/train.parquet
23
+ ---
24
+
25
+ # Agentic Prover RL
26
+
27
+ **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](https://github.com/GAIR-NLP/daVinci-prover) under **Lean 4.35.0**.
28
+
29
+ ## Sources
30
+
31
+ - **2,266 tasks:** our Lean formalizations of selected problems from NVIDIA's [nvidia/Nemotron-Math-Proofs-v3-RL](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v3-RL), whose source problems are licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
32
+ - **1,764 tasks:** our filtered oProof/Goedel pool. Original source identifiers are retained.
33
+
34
+ ## Load
35
+
36
+ ```python
37
+ from datasets import load_dataset
38
+
39
+ train = load_dataset("davinci-prover/agentic-prover-RL", split="train")
40
+ ```
41
+
42
+ ## Data format
43
+
44
+ | Field | Description |
45
+ |---|---|
46
+ | `task_id`, `task_type` | Unique task ID and `lean_theorem_proving`. |
47
+ | `theorem_statement` | Lean task source with the proof goal. |
48
+ | `target_workspace` | Pinned Mathlib repository and Lean toolchain. |
49
+ | `difficulty` | `easy` (1,313), `medium` (1,696), or `hard` (1,021). |
50
+ | `source`, `source_id`, `source_license` | Upstream provenance; license is null when absent from source metadata. |
51
+
52
+ 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`.
data/train.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7cf603bab81cab969bc574a4958ba0bce54009e8951949b342c24166d796aa70
3
+ size 1462943