README.md

---
license: apache-2.0
task_categories:
- question-answering
- text-generation
- summarization
language:
- zh
size_categories:
- 1K<n<10K
---

# PRGB Benchmark
PRGB (Placeholder RAG Benchmark) is a benchmark tool focused on evaluating document faithfulness and external knowledge utilization efficiency in Retrieval-Augmented Generation (RAG) systems. 
It comprehensively evaluates model performance through progressive dimensions such as multi-level filtering and cross-entity reasoning, using placeholders with noise-injected datasets to help researchers and developers analyze the performance of mainstream RAG models in complex scenarios.

For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to [PRGB GitHub](https://github.com/AQ-MedAI/PRGB) and PRGB Paper at [🤗 HuggingFace](https://huggingface.co/papers/2507.22927).

# Key Features
🎯 Multi-Model Support: Supports multiple large language models with local VLLM inference

📊 Standardized Evaluation: Provides unified evaluation metrics and processes

🔧 Flexible Configuration: Supports noise configuration, placeholder configuration, and other parameter adjustments

🌍 Multi-Language Support: Supports Chinese and English dataset evaluation

📈 Detailed Reports: Generates comprehensive evaluation results and score reports

## PRGB's Leaderboard

In our experiments, we uniformly set the following configurations:
`noise_config: '{"noise_doc_level1":4,"noise_doc_level2":4,"noise_doc_level3":1}'`, `num_iterations: 3`, and `shuffle: True`.

### Chinese Dataset Performance Comparison

The table below presents the performance of various state-of-the-art models on Chinese datasets, sorted by Overall score from highest to lowest. **Bold** values indicate the best experimental results, and ***italic bold*** values indicate the second-best experimental results.

| Models                        | Overall | Multi-Level Filter  | Composition         | Reasoning       |
| ----------------------------- | ------- | ------------------- | ------------------- | --------------- |
| `Gemini-2.5-pro-preview`    | 87.33   | **97.92**     | **94.20**     | 70.18           |
| `DeepSeek-R1-0528`          | 86.68   | 96.17               | ***93.69*** | **72.79** |
| `Claude-3.7-sonnet`         | 85.74   | ***97.62*** | 90.59               | **70.39** |
| `Gemini-2.5-flash-preview`  | 81.85   | 93.92               | 88.54               | 63.86           |
| `Qwen3-235B-A22B`           | 80.76   | 94.92               | 88.18               | 60.23           |
| `Qwen3-30B-A3B`             | 80.45   | 95.87               | 86.11               | 61.42           |
| `Deepseek-V3(241226)`       | 77.54   | 94.58               | 81.00               | 60.32           |
| `Qwen3-235B-A22B w/o think` | 75.20   | 91.50               | 79.67               | 57.14           |
| `Qwen-2.5-MAX`              | 74.43   | 93.25               | 78.28               | 55.37           |
| `Qwen3-30B-A3B w/o think`   | 71.05   | 91.08               | 72.22               | 54.76           |
| `Gemma3_27b`                | 70.24   | 73.09               | 92.21               | 50.24           |
| `Qwen3_32B`                 | 69.69   | 89.75               | 75.74               | 46.70           |
| `Hunyuan-80B-A13B`          | 68.84   | 93.50               | 68.94               | 50.64           |
| `GPT4.1`                    | 66.26   | 89.75               | 71.95               | 41.27           |
| `Qwen2.5_72B`               | 64.87   | 92.92               | 64.99               | 44.14           |
| `GPT4o-1120`                | 64.58   | 88.50               | 70.21               | 39.35           |
| `Gemma3_12b`                | 64.10   | 60.20               | 89.92               | 50.52           |
| `Qwen3_8B`                  | 63.04   | 86.87               | 67.49               | 39.47           |
| `Qwen3_32B w/o think`       | 60.73   | 59.53               | 89.50               | 41.30           |
| `Qwen2.5_32B`               | 58.76   | 92.00               | 51.33               | 44.60           |
| `Qwen2.5_14B`               | 55.94   | 89.42               | 52.69               | 35.87           |
| `Qwen2.5_7B`                | 49.31   | 83.29               | 47.47               | 26.92           |
| `Qwen3_8B w/o think`        | 50.02   | 47.83               | 83.96               | 28.17           |
| `Gemma3_4b`                 | 47.67   | 37.41               | 78.33               | 39.26           |



## Installation

### Requirements

- Python 3.7+
- CUDA (if using GPU inference)

### Installation Steps

1. Clone the repository

```bash
git clone https://github.com/Alipay-Med/PRGB.git
cd PRGB
```

2. Install dependencies

```bash
pip install -r requirements.txt
```

3. Verify installation

```bash
python test_imports.py
```

## Usage

### Verify Imports

Before running evaluations, it's recommended to verify that imports work correctly:

```bash
python test_imports.py
```

### Three Ways to Run Evaluation

#### Method 1: Using Makefile (Recommended)

If you only need to modify the model path, using Makefile is recommended

```bash
# View all available commands
make help

# Set environment variables and run evaluation
export EVAL_MODEL_PATH=/path/to/your/model
make eval

# Or set environment variables in one line
EVAL_MODEL_PATH=/path/to/your/model make eval

# Chinese evaluation (using data/zh.jsonl)
# Chinese evaluation with inference mode (using data/zh.jsonl)
EVAL_MODEL_PATH=/path/to/your/model make eval-ch-infer

# English evaluation (using data/en.jsonl)
# English evaluation with inference mode (using data/en.jsonl)
EVAL_MODEL_PATH=/path/to/your/model make eval-en-infer

# Test evaluation (no real model needed)
make eval-test

# Export error samples (requires evaluation result file path)
EVAL_RESULT_FILE=results/model_eval_result.jsonl make export-errors
```

#### Method 2: Using Shell Script

If you need to modify other parameters, using the shell is recommended.

```bash
# Run with default parameters (requires model path)
./run_eval.sh /path/to/your/model

# Pass all parameters
./run_eval.sh /path/to/your/model data/zh.jsonl Qwen3_infer ./results
```

#### Method 3: Using Python Command

```bash
# Basic usage
python eval.py \
    --model-name "Qwen3" \
    --model-path "/path/to/your/model" \
    --data-path "tests/test.jsonl" \
    --output-path "./results"

# Complete parameter example
python eval.py \
    --model-name "Qwen3" \
    --model-path "/path/to/your/model" \
    --data-path "your_data.jsonl" \
    --output-path "./results" \
    --batch-size 16 \
    --temperature 0.7 \
    --noise-config '{"noise_doc_level1":4,"noise_doc_level2":4,"noise_doc_level3":1}' \
    --custom_config "config/default_prompt_config.json" \
    --shuffle True \
    --num-iterations 3 \
    --verbose
```