Datasets:
text stringlengths 2.93k 989k |
|---|
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
Lean theorem-proving trajectory
Context:
A theorem-proving agent works in a Lean 4 and Mathlib workspace where the exact task is preloaded in `scratch/solution.lean`, and every Lean edit or write is checked automatically.
The trajectory succeeds only when that exact target is strictly verified without placeholders; au... |
"Lean theorem-proving trajectory\n\nContext:\nA theorem-proving agent works in a Lean 4 and Mathlib (...TRUNCATED) |
"Lean theorem-proving trajectory\n\nContext:\nA theorem-proving agent works in a Lean 4 and Mathlib (...TRUNCATED) |
"Lean theorem-proving trajectory\n\nContext:\nA theorem-proving agent works in a Lean 4 and Mathlib (...TRUNCATED) |
"Lean theorem-proving trajectory\n\nContext:\nA theorem-proving agent works in a Lean 4 and Mathlib (...TRUNCATED) |
Agentic Prover Midtrain
Lean theorem-proving text for mid-training. Each subset has a train split and a single text field. Original text is preserved in Parquet format, without tokenization or packing.
| Subset | Examples | Training tokens (B) | Content |
|---|---|---|---|
decompose |
4,662,378 | 8.435 | Theorem decomposition |
revision |
1,027,263 | 3.836 | Proof revision |
search |
1,961,667 | 3.447 | Parallel proof search and outcomes |
agent |
397,333 | 7.170 | Theorem-proving agent trajectories |
decompose-4.35-verified |
1,785,138 | 2.457 | Decompose examples passing Lean 4.35 verification |
revision-4.35-verified |
328,534 | 1.181 | Revision examples passing Lean 4.35 verification |
The four original subsets contain 8,048,641 examples and account for 22.888B training tokens. Informal data is excluded. Token counts use the original Qwen3.5 32K training preprocessing, including EOS and excluding replay and padding.
The two 4.35-verified subsets are filtered versions of decompose and revision, containing 2,113,672 examples in total. They overlap with the original subsets and should be used as alternatives when avoiding duplicate examples. Verification checks the final Lean proof against the original theorem using Lean 4.35.0; it does not verify natural-language explanations or initial proof attempts.
from datasets import load_dataset
train = load_dataset(
"davinci-prover/agentic-prover-midtrain",
"decompose-4.35-verified", # any subset listed above
split="train",
streaming=True,
)
Harness: daVinci-prover.
- Downloads last month
- 99