Commit ·
08d4c51
1
Parent(s): 474192c
bump to be compatible with Lean v4.34.0
Browse files- README.md +71 -28
- proofnet-verified.jsonl +0 -0
README.md
CHANGED
|
@@ -19,7 +19,7 @@ tags:
|
|
| 19 |
- HuggingFace: https://huggingface.co/datasets/marcusm117/ProofNet-Verified
|
| 20 |
- Visualization: https://marcusm117.github.io/ProofNet-Verified/
|
| 21 |
|
| 22 |
-
**ProofNet-Verified (ProofNet-V)** is an audited, corrected, and verified version of the [ProofNet](https://github.com/zhangir-azerbayev/ProofNet) dataset
|
| 23 |
|
| 24 |
[ProofNet](https://github.com/zhangir-azerbayev/ProofNet) is a benchmark for theorem proving and autoformalization of undergraduate-level mathematics. It originally consists of **371** data points, each consisting of a formal theorem statement in `Lean 3`, a natural language theorem statement, and a natural language proof. The problems are primarily drawn from popular undergraduate pure mathematics textbooks and cover topics such as real and complex analysis, linear algebra, abstract algebra, and topology. We intend for ProofNet to be a challenging benchmark that will drive progress in autoformalization and automatic theorem proving
|
| 25 |
|
|
@@ -36,9 +36,9 @@ Based on ProofNet and ProofNet#, we made the following modifications:
|
|
| 36 |
- Rudin|exercise_1_14
|
| 37 |
- Rudin|exercise_1_19
|
| 38 |
|
| 39 |
-
2. Upgrade the
|
| 40 |
|
| 41 |
-
3. Audit the benchmark for both errors in the ground-truth formalizations and ambiguity in the informal statements, identifying **
|
| 42 |
|
| 43 |
4. Confirm each error by proving the negation of incorrect formalizations, providing trivialized proofs of formalization that are provably true but unfaithful to the informal statement, or giving detailed reasoning for why a formalization is unfaithful or an informal statement is ambiguous.
|
| 44 |
|
|
@@ -72,30 +72,49 @@ If you find the **ProofNet-Verified** benchmark useful for your research, please
|
|
| 72 |
|
| 73 |
| | ProofNet | ProofNet# |
|
| 74 |
|---|---|---|
|
| 75 |
-
| Faithful |
|
| 76 |
-
| Unfaithful |
|
| 77 |
-
|
|
| 78 |
-
|
|
| 79 |
-
| **Total Errors** | **
|
| 80 |
| **Total** | **367** | **367** |
|
| 81 |
|
| 82 |
-
|
| 83 |
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 85 |
|---|---|---|
|
| 86 |
-
|
|
| 87 |
-
|
|
| 88 |
-
|
|
| 89 |
-
| Missing
|
| 90 |
-
|
|
| 91 |
-
|
|
| 92 |
-
|
|
| 93 |
-
|
|
|
|
|
|
|
|
|
|
|
| 94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 95 |
|
| 96 |
## Dataset Format
|
| 97 |
|
| 98 |
-
Each entry in [proofnet-verified.jsonl](
|
| 99 |
|
| 100 |
| Field | Description |
|
| 101 |
|---|---|
|
|
@@ -105,9 +124,9 @@ Each entry in [proofnet-verified.jsonl](data/proofnet-verified.jsonl) contains:
|
|
| 105 |
| `header` | Lean imports and `open` declarations |
|
| 106 |
| `helper` | Auxiliary definitions needed by some problems |
|
| 107 |
| `informal_stmt` | Natural language theorem statement |
|
| 108 |
-
| `formal_stmt` | Lean 4
|
| 109 |
| `informal_proof` | Natural language proof |
|
| 110 |
-
| `formal_proof` | [
|
| 111 |
|
| 112 |
### Example Entry
|
| 113 |
|
|
@@ -131,12 +150,24 @@ theorem Artin_exercise_2_2_9 {G : Type*} [Group G] {a b : G}
|
|
| 131 |
|
| 132 |
## Getting Started
|
| 133 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
### Prerequisites
|
| 135 |
|
| 136 |
-
- [Lean 4](https://lean-lang.org/lean4/doc/setup.html) (v4.
|
| 137 |
- [Lake](https://github.com/leanprover/lean4/tree/master/src/lake) (bundled with Lean)
|
| 138 |
- [uv](https://docs.astral.sh/uv/) (for Python scripts)
|
| 139 |
|
|
|
|
|
|
|
| 140 |
### Building
|
| 141 |
|
| 142 |
```bash
|
|
@@ -150,14 +181,26 @@ lake update
|
|
| 150 |
|
| 151 |
### Compiling All Statements
|
| 152 |
|
| 153 |
-
To type-check all 367 formal statements in parallel:
|
| 154 |
|
| 155 |
```bash
|
| 156 |
uv sync
|
| 157 |
uv run python scripts/verify_compilation.py
|
| 158 |
```
|
| 159 |
|
| 160 |
-
This runs `lake env lean` on each file with 48 parallel workers (configurable via `--jobs`).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 161 |
|
| 162 |
### Python Environment
|
| 163 |
|
|
@@ -175,16 +218,16 @@ uv run python scripts/verify_compilation.py
|
|
| 175 |
|
| 176 |
### For Theorem Proving Evaluation
|
| 177 |
|
| 178 |
-
The `proofnet-verified/` directory contains 367 Lean files, each with a single theorem
|
| 179 |
|
| 180 |
### For Autoformalization Evaluation
|
| 181 |
|
| 182 |
-
Each entry in `proofnet-verified.jsonl` pairs an `informal_stmt` with a `formal_stmt`. We recommend using symbolic checkers like [BEq+](https://github.com/augustepoiroux/LeanInteract/blob/main/examples/beq_plus.py) over LLM-as-a-Judge to compare the ground truth `formal_stmt` with the generated candidate formalizations.
|
| 183 |
|
| 184 |
### For Error Analysis
|
| 185 |
|
| 186 |
-
The `error_taxonomy/` directory provides detailed per-problem error classifications, useful for studying common failure modes in mathematical formalization (Lean syntax misunderstandings, Mathlib API misuse, missing premises, etc.).
|
| 187 |
|
| 188 |
## License
|
| 189 |
|
| 190 |
-
This project is licensed under the Apache License 2.0. See the [LICENSE](LICENSE) file for details.
|
|
|
|
| 19 |
- HuggingFace: https://huggingface.co/datasets/marcusm117/ProofNet-Verified
|
| 20 |
- Visualization: https://marcusm117.github.io/ProofNet-Verified/
|
| 21 |
|
| 22 |
+
**ProofNet-Verified (ProofNet-V)** is an audited, corrected, and verified version of the [ProofNet](https://github.com/zhangir-azerbayev/ProofNet) dataset. The latest project environment targets `Lean v4.34.0`; the verified Lean 4.28 snapshot remains available in [`data/archive/` on GitHub](https://github.com/marcusm117/ProofNet-Verified/tree/main/data/archive). A visualization of all errors, error classifications, and our corrections can be found at [https://marcusm117.github.io/ProofNet-Verified](https://marcusm117.github.io/ProofNet-Verified).
|
| 23 |
|
| 24 |
[ProofNet](https://github.com/zhangir-azerbayev/ProofNet) is a benchmark for theorem proving and autoformalization of undergraduate-level mathematics. It originally consists of **371** data points, each consisting of a formal theorem statement in `Lean 3`, a natural language theorem statement, and a natural language proof. The problems are primarily drawn from popular undergraduate pure mathematics textbooks and cover topics such as real and complex analysis, linear algebra, abstract algebra, and topology. We intend for ProofNet to be a challenging benchmark that will drive progress in autoformalization and automatic theorem proving
|
| 25 |
|
|
|
|
| 36 |
- Rudin|exercise_1_14
|
| 37 |
- Rudin|exercise_1_19
|
| 38 |
|
| 39 |
+
2. Upgrade the project environment to `Lean v4.34.0`, so that more automated proof search tools like [Grind](https://lean-lang.org/doc/reference/latest/The--grind--tactic/), [Canonical](https://github.com/chasenorman/CanonicalLean), and [Lean-SMT](https://github.com/ufmg-smite/lean-smt) are available.
|
| 40 |
|
| 41 |
+
3. Audit the benchmark for both errors in the ground-truth formalizations and ambiguity in the informal statements, identifying **205/367 (55.9%)** errors in **ProofNet** and **99/367 (27.0%)** errors in **ProofNet#**.
|
| 42 |
|
| 43 |
4. Confirm each error by proving the negation of incorrect formalizations, providing trivialized proofs of formalization that are provably true but unfaithful to the informal statement, or giving detailed reasoning for why a formalization is unfaithful or an informal statement is ambiguous.
|
| 44 |
|
|
|
|
| 72 |
|
| 73 |
| | ProofNet | ProofNet# |
|
| 74 |
|---|---|---|
|
| 75 |
+
| Faithful | 162 (44.1%) | 268 (73.0%) |
|
| 76 |
+
| Unfaithful | 195 (53.1%) | 90 (24.5%) |
|
| 77 |
+
| NL Source Ambiguous | 9 (2.5%) | 8 (2.2%) |
|
| 78 |
+
| NL Source Incorrect | 1 (0.3%) | 1 (0.3%) |
|
| 79 |
+
| **Total Errors** | **205 (55.9%)** | **99 (27.0%)** |
|
| 80 |
| **Total** | **367** | **367** |
|
| 81 |
|
| 82 |
+
The **Unfaithful** row counts the `stronger`, `weaker`, and `incomparable` faithfulness labels; ambiguous and incorrect natural-language sources are counted separately. Percentages use all 367 problems in each benchmark.
|
| 83 |
|
| 84 |
+
The audit tracks three dimensions: **Provability**, **Faithfulness**, and **Root Cause**. Faithfulness compares logical structure with the relevant mathematical formulas treated as opaque, rather than identifying statements because they are mathematically provable.
|
| 85 |
+
|
| 86 |
+
### Root Cause Breakdown
|
| 87 |
+
|
| 88 |
+
| Root Cause | ProofNet | ProofNet# |
|
| 89 |
|---|---|---|
|
| 90 |
+
| NL Source Ambiguous | 9 (2.5%) | 8 (2.2%) |
|
| 91 |
+
| NL Source Incorrect | 1 (0.3%) | 1 (0.3%) |
|
| 92 |
+
| Implicit Premise | 18 (4.9%) | 17 (4.6%) |
|
| 93 |
+
| Missing Premise | 31 (8.4%) | 6 (1.6%) |
|
| 94 |
+
| Adding Extra Conclusion | 7 (1.9%) | 3 (0.8%) |
|
| 95 |
+
| Misunderstanding of Informal Math | 95 (25.9%) | 30 (8.2%) |
|
| 96 |
+
| Misunderstanding of Mathlib | 28 (7.6%) | 24 (6.5%) |
|
| 97 |
+
| Misunderstanding of Lean | 16 (4.4%) | 10 (2.7%) |
|
| 98 |
+
| **Total Errors** | **205 (55.9%)** | **99 (27.0%)** |
|
| 99 |
+
|
| 100 |
+
These are the eight root-cause categories. Faithful formalizations use `none`, meaning **no formalization error / not applicable**, not an unknown cause or a ninth error category.
|
| 101 |
|
| 102 |
+
## Dataset Files
|
| 103 |
+
|
| 104 |
+
| Resource | Location |
|
| 105 |
+
|---|---|
|
| 106 |
+
| Latest JSONL dataset on Hugging Face | [`proofnet-verified.jsonl`](proofnet-verified.jsonl) |
|
| 107 |
+
| Canonical latest JSONL dataset on GitHub | [`data/proofnet-verified-latest.jsonl`](https://github.com/marcusm117/ProofNet-Verified/blob/main/data/proofnet-verified-latest.jsonl) |
|
| 108 |
+
| Latest Lean statements | [`proofnet-verified-latest/`](https://github.com/marcusm117/ProofNet-Verified/tree/main/proofnet-verified-latest) |
|
| 109 |
+
| Archived Lean 4.28 statements | [`data/archive/proofnet-verified-4.28/`](https://github.com/marcusm117/ProofNet-Verified/tree/main/data/archive/proofnet-verified-4.28) |
|
| 110 |
+
| Archived Lean 4.28 JSONL dataset | [`data/archive/proofnet-verified-4.28.jsonl`](https://github.com/marcusm117/ProofNet-Verified/blob/main/data/archive/proofnet-verified-4.28.jsonl) |
|
| 111 |
+
| Archived Lean 4.28 proofs | [`data/archive/proofnet-verified_gt-4.28.zip`](https://github.com/marcusm117/ProofNet-Verified/blob/main/data/archive/proofnet-verified_gt-4.28.zip) |
|
| 112 |
+
|
| 113 |
+
The Hugging Face JSONL is a byte-for-byte copy of the canonical latest GitHub export, including its existing proof metadata. Its filename remains `proofnet-verified.jsonl` for compatibility.
|
| 114 |
|
| 115 |
## Dataset Format
|
| 116 |
|
| 117 |
+
Each entry in [proofnet-verified.jsonl](proofnet-verified.jsonl) contains:
|
| 118 |
|
| 119 |
| Field | Description |
|
| 120 |
|---|---|
|
|
|
|
| 124 |
| `header` | Lean imports and `open` declarations |
|
| 125 |
| `helper` | Auxiliary definitions needed by some problems |
|
| 126 |
| `informal_stmt` | Natural language theorem statement |
|
| 127 |
+
| `formal_stmt` | Lean 4 theorem declaration |
|
| 128 |
| `informal_proof` | Natural language proof |
|
| 129 |
+
| `formal_proof` | Placeholder preserved from the canonical export; the Lean 4.28 proof snapshot is [available as a ZIP](https://github.com/marcusm117/ProofNet-Verified/blob/main/data/archive/proofnet-verified_gt-4.28.zip) |
|
| 130 |
|
| 131 |
### Example Entry
|
| 132 |
|
|
|
|
| 150 |
|
| 151 |
## Getting Started
|
| 152 |
|
| 153 |
+
### Loading from Hugging Face
|
| 154 |
+
|
| 155 |
+
```python
|
| 156 |
+
from datasets import load_dataset
|
| 157 |
+
|
| 158 |
+
dataset = load_dataset("marcusm117/ProofNet-Verified")
|
| 159 |
+
```
|
| 160 |
+
|
| 161 |
+
The Lean statements, build configuration, and Python scripts live in the [GitHub repository](https://github.com/marcusm117/ProofNet-Verified), not this Hugging Face repository. Run the build, compilation, and regeneration commands below from the GitHub checkout.
|
| 162 |
+
|
| 163 |
### Prerequisites
|
| 164 |
|
| 165 |
+
- [Lean 4](https://lean-lang.org/lean4/doc/setup.html) (v4.34.0, managed via `elan`)
|
| 166 |
- [Lake](https://github.com/leanprover/lean4/tree/master/src/lake) (bundled with Lean)
|
| 167 |
- [uv](https://docs.astral.sh/uv/) (for Python scripts)
|
| 168 |
|
| 169 |
+
The latest statements use Mathlib and Canonical `v4.34.0`, with Lean-SMT pinned to [`2956f2966ff443b27a86fe2a4fb82dc56efafc51`](https://github.com/ufmg-smite/lean-smt/commit/2956f2966ff443b27a86fe2a4fb82dc56efafc51). The archived and original Lean 4.28 datasets remain unchanged.
|
| 170 |
+
|
| 171 |
### Building
|
| 172 |
|
| 173 |
```bash
|
|
|
|
| 181 |
|
| 182 |
### Compiling All Statements
|
| 183 |
|
| 184 |
+
To type-check all 367 formal statements in `proofnet-verified-latest/` in parallel:
|
| 185 |
|
| 186 |
```bash
|
| 187 |
uv sync
|
| 188 |
uv run python scripts/verify_compilation.py
|
| 189 |
```
|
| 190 |
|
| 191 |
+
This runs `lake env lean` on each file with 48 parallel workers (configurable via `--jobs`, e.g. `--jobs 64`). The latest statements use the repository's Lean 4.34 environment. For the archived statements (`--dir data/archive/proofnet-verified-4.28`) or original formalizations (`--dir data/proofnet/proofnet-4.28`), also pass `--cwd /path/to/lean-4.28-project` pointing to a Lake project with Lean and Mathlib 4.28; overriding `--lean-version` alone does not change the dependency versions.
|
| 192 |
+
|
| 193 |
+
### Regenerating JSONL
|
| 194 |
+
|
| 195 |
+
Run these commands from the GitHub repository root to regenerate the latest dataset, the archived Lean 4.28 dataset, and the original ProofNet dataset:
|
| 196 |
+
|
| 197 |
+
```bash
|
| 198 |
+
uv run python scripts/generate_jsonl.py --source-dir proofnet-verified-latest --output data/proofnet-verified-latest.jsonl
|
| 199 |
+
uv run python scripts/generate_jsonl.py --source-dir data/archive/proofnet-verified-4.28 --output data/archive/proofnet-verified-4.28.jsonl
|
| 200 |
+
uv run python scripts/generate_jsonl.py --source-dir data/proofnet/proofnet-4.28 --output data/proofnet/proofnet-4.28.jsonl
|
| 201 |
+
```
|
| 202 |
+
|
| 203 |
+
The exporter preserves `informal_proof` and `formal_proof` from the existing output. For a new output file, pass `--metadata` with an existing corresponding JSONL dataset. Add `--check` to validate an export without rewriting it.
|
| 204 |
|
| 205 |
### Python Environment
|
| 206 |
|
|
|
|
| 218 |
|
| 219 |
### For Theorem Proving Evaluation
|
| 220 |
|
| 221 |
+
The [`proofnet-verified-latest/` directory on GitHub](https://github.com/marcusm117/ProofNet-Verified/tree/main/proofnet-verified-latest) contains 367 Lean files, each with a single theorem declaration. Replace `sorry` with a valid proof for entries with an admitted proof. We recommend using [Lean Comparator](https://github.com/leanprover/comparator) as a rigorous proof checker.
|
| 222 |
|
| 223 |
### For Autoformalization Evaluation
|
| 224 |
|
| 225 |
+
Each entry in [`proofnet-verified.jsonl`](proofnet-verified.jsonl) pairs an `informal_stmt` with a `formal_stmt`. We recommend using symbolic checkers like [BEq+](https://github.com/augustepoiroux/LeanInteract/blob/main/examples/beq_plus.py) over LLM-as-a-Judge to compare the ground truth `formal_stmt` with the generated candidate formalizations.
|
| 226 |
|
| 227 |
### For Error Analysis
|
| 228 |
|
| 229 |
+
The [`error_taxonomy/` directory on GitHub](https://github.com/marcusm117/ProofNet-Verified/tree/main/error_taxonomy) provides detailed per-problem error classifications, useful for studying common failure modes in mathematical formalization (Lean syntax misunderstandings, Mathlib API misuse, missing premises, etc.).
|
| 230 |
|
| 231 |
## License
|
| 232 |
|
| 233 |
+
This project is licensed under the Apache License 2.0. See the [LICENSE](LICENSE) file for details.
|
proofnet-verified.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|