2026-07-08 08:57:50 +00:00

133 lines
3.7 KiB
Markdown

# EQ-Bench
## Overview
EQ-Bench is a benchmark for evaluating language models on emotional intelligence tasks. It assesses the ability to predict likely emotional responses of characters in dialogues by rating the intensity of possible emotional reactions.
## Task Description
- **Task Type**: Emotional Intelligence Assessment
- **Input**: Dialogue scenario with characters
- **Output**: Emotion intensity ratings in specific format
- **Domains**: Emotional understanding, social cognition
## Key Features
- Tests ability to predict emotional responses in conversations
- Requires rating intensity of multiple possible emotions
- Uses official EQ-Bench v2 scoring algorithm
- Scoring includes sigmoid scaling for small differences
- Adjustment constant ensures random answers score 0
## Evaluation Notes
- Default evaluation uses the **validation** split
- Primary metric: **EQ-Bench Score** (0-100 scale, reported as 0-1)
- Uses zero-shot evaluation (no few-shot examples)
- Responses must include emotion ratings in specific JSON-like format
- Official algorithm from [Paper](https://arxiv.org/abs/2312.06281) | [Homepage](https://eqbench.com/)
## Properties
| Property | Value |
|----------|-------|
| **Benchmark Name** | `eq_bench` |
| **Dataset ID** | [evalscope/EQ-Bench](https://modelscope.cn/datasets/evalscope/EQ-Bench/summary) |
| **Paper** | N/A |
| **Tags** | `InstructionFollowing` |
| **Metrics** | `eq_bench_score` |
| **Default Shots** | 0-shot |
| **Evaluation Split** | `validation` |
## Data Statistics
| Metric | Value |
|--------|-------|
| Total Samples | 171 |
| Prompt Length (Mean) | 1550.02 chars |
| Prompt Length (Min/Max) | 922 / 3737 chars |
## Sample Example
**Subset**: `default`
```json
{
"input": [
{
"id": "97129bc9",
"content": "Your task is to predict the likely emotional responses of a character in this dialogue:\n\nRobert: Claudia, you've always been the idealist. But let's be practical for once, shall we?\nClaudia: Practicality, according to you, means bulldozing ev ... [TRUNCATED] ... ary:\n\nRemorseful: <score>\nIndifferent: <score>\nAffectionate: <score>\nAnnoyed: <score>\n\n\n[End of answer]\n\nRemember: zero is a valid score, meaning they are likely not feeling that emotion. You must score at least one emotion > 0.\n\nYour answer:"
}
],
"target": "{'emotion1': 'Remorseful', 'emotion2': 'Indifferent', 'emotion3': 'Affectionate', 'emotion4': 'Annoyed', 'emotion1_score': 2, 'emotion2_score': 3, 'emotion3_score': 0, 'emotion4_score': 5}",
"id": 0,
"group_id": 0,
"metadata": {
"reference_answer": {
"emotion1": "Remorseful",
"emotion2": "Indifferent",
"emotion3": "Affectionate",
"emotion4": "Annoyed",
"emotion1_score": 2,
"emotion2_score": 3,
"emotion3_score": 0,
"emotion4_score": 5
},
"reference_answer_fullscale": {
"emotion1": "Remorseful",
"emotion2": "Indifferent",
"emotion3": "Affectionate",
"emotion4": "Annoyed",
"emotion1_score": 0,
"emotion2_score": "6",
"emotion3_score": 0,
"emotion4_score": "7"
}
}
}
```
*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 eq_bench \
--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=['eq_bench'],
limit=10, # Remove this line for formal evaluation
)
run_task(task_cfg=task_cfg)
```