Datasets:

ArXiv:
File size: 4,974 Bytes
9f3bc09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
# Training Qwen2.5-3B-Instruct for Evaluation Agent with CoT Reasoning

This repository contains scripts and configurations for training Qwen2.5-3B-Instruct model on evaluation agent data with Chain-of-Thought (CoT) reasoning format.

## Overview

The training pipeline processes evaluation results from:
- **VBench**: Video quality evaluation results
- **T2I-CompBench**: Text-to-image composition evaluation results  
- **Open Domain**: Open-ended query evaluation results

All results are in CoT (Chain-of-Thought) reasoning format from proprietary models.

## Dataset Preparation

### 1. Data Cleaning and Conversion

Run the data cleaning script to convert raw evaluation results into LLaMA-Factory format:

```bash
python clean_and_convert_data.py
```

This script:
- Processes JSON files from `ea-data/agent/` subdirectories
- Converts CoT-style evaluation results into instruction-response pairs
- Outputs to `LLaMA-Factory/data/evaluation_agent_cot_dataset.json`
- Updates `LLaMA-Factory/data/dataset_info.json` with dataset metadata

### Dataset Statistics
- Total training examples: ~860 (from initial processing)
- Format: Alpaca-style (instruction, input, output)

## Training Configurations

### 1. LoRA Fine-tuning (Recommended)

**Configuration:** `train_qwen2.5_eval_agent.yaml`

Key parameters:
- Model: Qwen/Qwen2.5-3B-Instruct
- Method: LoRA (rank=16, alpha=32)
- Batch size: 2 per device × 4 gradient accumulation
- Learning rate: 5e-5 with cosine scheduler
- Epochs: 3
- Memory requirement: ~16GB VRAM

### 2. Full Fine-tuning

**Configuration:** `train_qwen2.5_eval_agent_full.yaml`

Key parameters:
- Model: Qwen/Qwen2.5-3B-Instruct
- Method: Full fine-tuning with DeepSpeed
- Gradient checkpointing enabled
- Memory requirement: ~32GB+ VRAM

## Training Execution

### Quick Start

```bash
# Make script executable
chmod +x train_qwen2.5_eval_agent.sh

# Run training
./train_qwen2.5_eval_agent.sh
```

### Manual Training

```bash
cd LLaMA-Factory
llamafactory-cli train ../train_qwen2.5_eval_agent.yaml
```

### Distributed Training

For multi-GPU training:

```bash
CUDA_VISIBLE_DEVICES=0,1,2,3 \
torchrun --nproc_per_node 4 \
--master_port 29500 \
src/train.py ../train_qwen2.5_eval_agent.yaml
```

## Inference

After training, run inference with:

```bash
llamafactory-cli chat ../inference_qwen2.5_eval_agent.yaml
```

Or use the API:

```bash
llamafactory-cli api ../inference_qwen2.5_eval_agent.yaml
```

## Model Merging

To merge LoRA weights with base model:

```bash
llamafactory-cli export \
    --model_name_or_path Qwen/Qwen2.5-3B-Instruct \
    --adapter_name_or_path saves/qwen2.5-3b/lora/eval_agent_cot \
    --template qwen \
    --finetuning_type lora \
    --export_dir models/qwen2.5-3b-eval-agent-merged \
    --export_size 4 \
    --export_legacy_format false
```

## Monitoring Training

### TensorBoard

```bash
tensorboard --logdir saves/qwen2.5-3b/lora/eval_agent_cot
```

### Loss Plots

Training loss plots are automatically saved to the output directory.

## Evaluation

The model will be evaluated on:
- CoT reasoning quality
- Evaluation accuracy
- Response coherence
- Format consistency

## Directory Structure

```
evaluation_agent_dev/
├── ea-data/agent/           # Raw evaluation data
│   ├── vbench_results/
│   ├── t2i_results/
│   └── open_results/
├── LLaMA-Factory/           # Training framework
│   └── data/
│       ├── evaluation_agent_cot_dataset.json  # Processed dataset
│       └── dataset_info.json
├── clean_and_convert_data.py    # Data processing script
├── train_qwen2.5_eval_agent.yaml    # LoRA training config
├── train_qwen2.5_eval_agent_full.yaml   # Full training config
├── inference_qwen2.5_eval_agent.yaml    # Inference config
└── train_qwen2.5_eval_agent.sh      # Training script
```

## Requirements

- Python 3.9+
- PyTorch 2.0+
- CUDA 11.6+
- LLaMA-Factory (installed)
- 16GB+ VRAM for LoRA, 32GB+ for full fine-tuning

## Tips

1. **Memory Management**: Use gradient checkpointing and DeepSpeed for larger batch sizes
2. **Learning Rate**: Start with 5e-5 for LoRA, 2e-5 for full fine-tuning
3. **Data Quality**: Review generated dataset for quality before training
4. **Checkpointing**: Save checkpoints frequently (every 200 steps)
5. **Mixed Precision**: Use bf16 for faster training and lower memory usage

## Troubleshooting

- **OOM Errors**: Reduce batch size or enable gradient checkpointing
- **Slow Training**: Enable Flash Attention 2 if available
- **Poor Results**: Increase training epochs or adjust learning rate
- **Data Issues**: Check JSON parsing in data cleaning script

## Next Steps

1. Expand dataset with more evaluation examples
2. Implement custom evaluation metrics
3. Fine-tune on specific evaluation dimensions
4. Deploy model for production use

## License

Follow the licenses of:
- Qwen2.5 model
- LLaMA-Factory framework
- Original evaluation datasets