| # EmailResponder-MCP Evaluation Tools |
|
|
| Evaluation scripts for analyzing model performance on the EmailResponder-MCP task. |
|
|
| ## Scripts |
|
|
| ### evaluate_success.py |
| Calculates success rates for each model based on execution logs. |
| |
| ```bash |
| python evaluate_success.py |
| ``` |
| |
| Outputs: |
| - `success-finish_detailed_results.json` - Detailed results per session |
| - `success-finish_rate.csv` - Summary statistics |
| |
| ### evaluate_trajectory.py |
| Evaluates trajectory metrics comparing actual execution paths against reference trajectory. |
| |
| ```bash |
| python evaluate_trajectory.py --config reference_trajectory.yaml --output evaluation_results.csv |
| ``` |
| |
| Metrics: |
| - Exact match |
| - In-order match |
| - Any-order match |
| - Precision / Recall |
| - Single-tool use |
| - Path diversity (unique_path_ratio, path_entropy) |
|
|
| ### analyze_retry_patterns.py |
| Analyzes retry patterns and error distributions across models. |
|
|
| ```bash |
| python analyze_retry_patterns.py |
| ``` |
|
|
| Outputs: |
| - `retry_analysis.json` - Detailed retry analysis |
| - `retry_summary.csv` - Retry statistics summary |
| - `error_by_agent.csv` - Error counts by agent |
|
|
| ## Configuration |
|
|
| Edit `reference_trajectory.yaml` to customize: |
| - Reference trajectory (ground truth) |
| - Target tools for single-tool use metric |
| - Models to evaluate |
| - Extraction types (SPAN, Chain, AGENT, LLM, Tool) |
|
|
| ## Requirements |
|
|
| - Python 3.7+ |
| - pandas |
| - pyyaml |
|
|
