Update README.md

README.md

@@ -6,3 +6,78 @@ configs:
   - split: test
     path: "dataset.parquet"
 ---
+# InvaR1ant Benchmark Dataset Card
+
+## Overview
+
+The InvaR1ant benchmark is a dataset designed to test the ability of language models to generalise invariant logical reasoning across different input sizes. Each example contains:
+
+- A **question** with multiple examples of constraints for different input sizes N
+- A target input size to predict constraints for
+- An **answer** with the correct constraint for the target size
+- Metadata including difficulty tier
+
+## Dataset Statistics
+
+### General Statistics
+- **Total samples**: 671
+- **Format**: Parquet file
+
+### Token Lengths (Qwen2.5-3B tokenizer)
+| Statistic | Questions | Answers |
+|-----------|-----------|---------|
+| Min       | 117       | 27      |
+| Max       | 4,086     | 14,476  |
+| Mean      | 1,684     | 994     |
+| Median    | 1,521     | 493     |
+| Total     | 1,130,153 | 666,709 |
+
+### Difficulty Distribution
+| Tier    | Count | Percentage | Avg Question Tokens | Avg Answer Tokens |
+|---------|-------|------------|---------------------|-------------------|
+| Small   | 333   | 49.6%      | 1,797               | 811               |
+| Medium  | 333   | 49.6%      | 1,580               | 1,161             |
+| Large   | 5     | 0.7%       | 1,107               | 2,022             |
+
+*Note: Difficulty tier is based on the "jump" between the largest example N and the target N*
+
+### Target N Distribution
+- **Range**: 5 to 30
+- **Mean**: 19.9
+- **Median**: 20
+- **Unique values**: 25 different values [5, 7, 8, 9, ..., 30]
+
+### Examples Per Question
+- **Range**: 3 to 18 examples
+- **Mean**: 6.7
+- **Median**: 6
+- **Most common**: 3 examples (168 samples)
+
+## Dataset Structure
+
+### Input Format
+Each question follows this template:
+```
+Given the following examples of constraints for increasing input sizes:
+N=3: (assert (and (and (and ...)))
+N=4: (assert (and (and (and ...)))
+...
+What is the constraint for N={target}?
+```
+
+### Output Format
+Answers are SMT constraints for the target N value, typically in the form:
+```
+(assert (and (and (and ...))))
+```
+
+## Limitations
+
+- Large tier samples are underrepresented (only 0.7% of the dataset)
+- Question token length (up to 4,086 tokens) may exceed context windows of some models
+- The dataset focuses specifically on SMT constraint generalization and may not reflect broader reasoning capabilities
+- These answers were generated from a symbolic execution tool, therefore the constraints may not be in the simplest format. You should verify equivalence with a theorem prover like Z3
+
+---
+
+*Dataset created using the Qwen/Qwen2.5-3B tokenizer for length calculations*