| ## Overview |
|
|
| The evaluation system consists of three main components: |
|
|
| 1. **`run_generation_hf.py`**: Runs inference for individual datasets |
| 2. **`get_scores.py`**: Modular evaluation script that calculates scores |
| 3. **`run_all_evaluation.py`**: Comprehensive wrapper for running full pipelines |
| |
| ## Inference Step Customization |
| |
| **The inference step must be modified by users based on their specific model requirements.** |
| |
| As the model landscape continuously expands and evolves, the inference scripts provided are **reference implementations** that need to be adapted for your use case. Different models have different: |
| - Loading mechanisms |
| - Tokenization requirements |
| - Generation parameters |
| - API interfaces |
| - Memory requirements |
| |
| ### Sample Inference Implementations |
| |
| We provide two sample inference scripts - `run_generation_hf.py` and `run_generation_vllm.py` |
| |
| ### How to Customize |
| |
| 1. **Choose or create an inference script** that matches your model's requirements |
| 2. **Modify the model loading** section to work with your specific model |
| 3. **Adjust generation parameters** (temperature, top_p, max_tokens, etc.) |
| 4. **Update the prompt formatting** if your model uses a different template |
| |
| For comprehensive examples of different usage patterns, see **[`example_usage.sh`](./example_usage.sh)**, which includes: |
| - Full pipeline execution |
| - Inference-only runs |
| - Evaluation-only runs |
| |
| **After generating predictions, the evaluation step (`get_scores.py`) remains the same across all models.** |
