WARP-benchmark / README.md
dannkoh's picture
Update README.md
0d40985 verified
|
Raw
History Blame Contribute Delete
3.13 kB
---
license: mit
configs:
- config_name: default
data_files:
- split: test
path: "dataset.parquet"
---
# WARP-benchmark
The benchmark to test logical reasoning and pattern generalisation in large language models through formal SMT constraints.
## Overview
We propose a benchmark designed to evaluate a language model's ability to generalise worst-case path constraints for all input sizes. Each example asks a model to generate a formal constraint for a specific target size after being shown examples of constraints for smaller input sizes.
## Github Repository
You can find the link here at: [WARP-benchmark](https://github.com/dannkoh/warp-benchmark)
## Dataset Statistics
### Basic Statistics
- **Total samples**: 671
- **Format**: Parquet file
- **Difficulty tiers**:
- Small: 333 samples (49.6%)
- Medium: 333 samples (49.6%)
- Large: 5 samples (0.7%)
### Token Statistics (using Qwen2.5-3B tokenizer)
| | Questions | Answers |
|----------------|-----------|---------|
| Minimum | 117 | 27 |
| Maximum | 4,086 | 14,476 |
| Mean | 1,684 | 994 |
| Median | 1,521 | 493 |
| Total tokens | 1,130,153 | 666,709 |
### Target N Distribution
- Range: 5 to 30
- Mean: 19.9
- Median: 20
- Unique values: 25 different values
### Examples Per Question
- Range: 3 to 18 examples
- Mean: 6.7
- Median: 6
- Most common: 3 examples (168 questions)
## Dataset Structure
Each example in the dataset follows this structure:
- **Question**: A prompt showing examples of constraints for smaller input sizes N, asking for constraints at target N
- **Answer**: The formal constraint for the target size
- **Constants**: Variable declarations for the target size
- **Tier**: Difficulty level ("small", "medium", or "large")
Example question format:
```
Given the following examples of constraints for increasing input sizes:
N=3: (assert (and (and ...)))
N=4: (assert (and (and ...)))
...
What is the constraint for N={target}?
```
## Dataset Creation
The dataset is generated from SMT2 files using the following process:
1. **Discovery**: Finds SMT2 files following the naming pattern `{prefix}.{problem}_{number}.smt2`
2. **Combination Generation**: For each problem and target N value, generates combinations of smaller N values to use as examples
3. **Prompt Construction**: Creates prompts with selected examples and targets, ensuring they fit within token limits
4. **Difficulty Assignment**: Assigns difficulty tiers based on the "jump" between example N and target N:
- Small: ≤ 5
- Medium: 6-15
- Large: > 15
5. **Balanced Sampling**: Uses reservoir sampling to balance the dataset across difficulty tiers
## Dataset Profiling
Use the included profiler to analyze the dataset:
```bash
python profiler.py dataset.parquet --visualize
```
This will generate comprehensive statistics and visualizations about token lengths, difficulty distribution, and other metrics.
## Dependencies
- pandas
- pyarrow
- transformers (Qwen2.5-3B tokenizer)
- tqdm
- matplotlib (for visualization)
- seaborn (for visualization)