115 lines
3.0 KiB
Markdown
115 lines
3.0 KiB
Markdown
# AlpacaEval2.0
|
|
|
|
|
|
## Overview
|
|
|
|
AlpacaEval 2.0 is an evaluation framework for instruction-following language models that uses an LLM judge to compare model outputs against a strong baseline. It provides win-rate metrics reflecting human preferences.
|
|
|
|
## Task Description
|
|
|
|
- **Task Type**: Instruction-Following Evaluation (Pairwise Comparison)
|
|
- **Input**: User instruction/question
|
|
- **Output**: Model response compared against GPT-4 Turbo baseline
|
|
- **Metric**: Win rate against baseline model
|
|
|
|
## Key Features
|
|
|
|
- Auto-annotator for scalable evaluation
|
|
- Compares against GPT-4 Turbo baseline outputs
|
|
- High correlation with human preferences
|
|
- Cost-effective evaluation method
|
|
- Tests general instruction-following capabilities
|
|
|
|
## Evaluation Notes
|
|
|
|
- Default configuration uses **0-shot** evaluation
|
|
- Uses LLM judge (default: gpt-4-1106-preview)
|
|
- Baseline model: gpt-4-turbo outputs
|
|
- Reports win rate metric
|
|
- Note: Length-controlled win rate not currently supported
|
|
|
|
|
|
## Properties
|
|
|
|
| Property | Value |
|
|
|----------|-------|
|
|
| **Benchmark Name** | `alpaca_eval` |
|
|
| **Dataset ID** | [AI-ModelScope/alpaca_eval](https://modelscope.cn/datasets/AI-ModelScope/alpaca_eval/summary) |
|
|
| **Paper** | N/A |
|
|
| **Tags** | `Arena`, `InstructionFollowing` |
|
|
| **Metrics** | `winrate` |
|
|
| **Default Shots** | 0-shot |
|
|
| **Evaluation Split** | `eval` |
|
|
|
|
|
|
## Data Statistics
|
|
|
|
| Metric | Value |
|
|
|--------|-------|
|
|
| Total Samples | 805 |
|
|
| Prompt Length (Mean) | 164.92 chars |
|
|
| Prompt Length (Min/Max) | 12 / 1917 chars |
|
|
|
|
## Sample Example
|
|
|
|
**Subset**: `alpaca_eval_gpt4_baseline`
|
|
|
|
```json
|
|
{
|
|
"input": [
|
|
{
|
|
"id": "95236545",
|
|
"content": "What are the names of some famous actors that started their careers on Broadway?"
|
|
}
|
|
],
|
|
"target": "Several famous actors started their careers on Broadway before making it big in film and television. Here are a few notable examples:\n\n1. Sarah Jessica Parker - Before she was Carrie Bradshaw on \"Sex and the City,\" Sarah Jessica Parker was a ... [TRUNCATED] ... f the many performers who have transitioned from the Broadway stage to broader fame in the entertainment industry. Broadway often serves as a proving ground for talent, and many actors continue to return to the stage throughout their careers.",
|
|
"id": 0,
|
|
"group_id": 0,
|
|
"metadata": {
|
|
"generator": "gpt4_1106_preview",
|
|
"dataset": "helpful_base"
|
|
}
|
|
}
|
|
```
|
|
|
|
*Note: Some content was truncated for display.*
|
|
|
|
## Prompt Template
|
|
|
|
**Prompt Template:**
|
|
```text
|
|
{question}
|
|
```
|
|
|
|
## Usage
|
|
|
|
### Using CLI
|
|
|
|
```bash
|
|
evalscope eval \
|
|
--model YOUR_MODEL \
|
|
--api-url OPENAI_API_COMPAT_URL \
|
|
--api-key EMPTY_TOKEN \
|
|
--datasets alpaca_eval \
|
|
--limit 10 # Remove this line for formal evaluation
|
|
```
|
|
|
|
### Using Python
|
|
|
|
```python
|
|
from evalscope import run_task
|
|
from evalscope.config import TaskConfig
|
|
|
|
task_cfg = TaskConfig(
|
|
model='YOUR_MODEL',
|
|
api_url='OPENAI_API_COMPAT_URL',
|
|
api_key='EMPTY_TOKEN',
|
|
datasets=['alpaca_eval'],
|
|
limit=10, # Remove this line for formal evaluation
|
|
)
|
|
|
|
run_task(task_cfg=task_cfg)
|
|
```
|
|
|
|
|