| # Dataset Card for TMBench | |
| ## Dataset Summary | |
| TMBench is the benchmark proposed in the paper [Computational Reasoning of Large Language Models](https://arxiv.org/pdf/2504.20771?). It is designed to access the ability of Computational Reasoning of LLMs. | |
| The benchmark includes four variants: | |
| - **TMBench** (Latin): uses standard Latin alphabet symbols | |
| - **TMBenchGreek**: includes Greek symbols | |
| - **TMBenchNumber**: includes numerical digits | |
| - **TMBenchSpecial**: includes special characters | |
| Each sample contains: | |
| - An initial string | |
| - Number of deleted characters per step | |
| - A set of production rules (symbol → string) | |
| - Step-wise results | |
| ## Dataset Structure | |
| ### Data Fields | |
| Each example contains the following fields: | |
| - `id` (string): Unique ID of the sample | |
| - `init_str` (string): The initial string before rewriting | |
| - `rule` (dict): Rewriting rules (symbol → replacement string) | |
| - `delete_count` (int): Number of characters deleted at each step | |
| - `step_results` (list[string]): The string after each rewriting step | |
| ### Example | |
| ```json | |
| { | |
| "id": "001", | |
| "init_str": "abba", | |
| "rule": { | |
| "a": "ab", | |
| "b": "ba" | |
| }, | |
| "delete_count": 2, | |
| "step_results": ["abba", "baab", "abba"] | |
| } | |