README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- text-generation
- evaluation-agent
- cot-reasoning
- checkpoint
- qwen2.5
- video-assessment
- image-assessment
library_name: transformers
pipeline_tag: text-generation
---

# ea-dev-final

This is checkpoint **final** (step 471) from fine-tuning [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) for evaluation agent tasks.

## Checkpoint Details

- **Checkpoint**: final
- **Global Step**: 471
- **Epoch**: 3.00
- **Training Loss**: 0.8296
- **Learning Rate**: unknown
- **Base Model**: Qwen2.5-3B-Instruct
- **Task**: Multi-modal quality assessment with CoT reasoning

## Model Description

This checkpoint is from training an evaluation agent that can assess:
- **Video Quality**: Temporal consistency, motion smoothness, object consistency (VBench)
- **Image Quality**: Aesthetic quality, semantic alignment, visual fidelity (T2I-CompBench)
- **Open-ended Evaluation**: Custom quality assessment tasks

The model uses Chain-of-Thought (CoT) reasoning to provide detailed explanations for its evaluations.

## Files Included

This checkpoint contains:
- **Model Weights**: `model*.safetensors` - The actual model parameters  
- **Tokenizer**: Complete tokenizer configuration and vocabulary
- **Configuration**: Model and generation configuration files

**Note**: This checkpoint contains only inference files (no optimizer states).

## Usage

### For Inference
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the checkpoint
model = AutoModelForCausalLM.from_pretrained(
    "ea-dev-final",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ea-dev-final")

# Example evaluation prompt
prompt = """Please evaluate the quality of this video based on the following criteria:
1. Visual quality and clarity
2. Temporal consistency
3. Motion smoothness

Video description: A person walking through a park with trees swaying in the wind.

Let me think step by step:"""

inputs = tokenizer(prompt, return_tensors="pt")
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_length=512,
        do_sample=True,
        temperature=0.7,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

### Resume Training (if optimizer states included)
```bash
# Use with LLaMA-Factory
llamafactory-cli train \
    --stage sft \
    --model_name_or_path ea-dev-final \
    --resume_from_checkpoint ea-dev-final
```

## Training Progress

This checkpoint represents an intermediate state in the training process:
- **Steps Completed**: 471
- **Epochs**: 3.00
- **Current Loss**: 0.8296

## Related Models

This checkpoint is part of a series. Other checkpoints from the same training run:
- Look for repositories with pattern: `ea-dev-checkpoint-*`
- Final model: `ea-dev-final`

## License

This model checkpoint is released under the Apache 2.0 license.

## Citation

If you use this checkpoint, please cite:
```bibtex
@misc{eval-agent-qwen2.5-checkpoint-471,
  title={Evaluation Agent Qwen2.5 Checkpoint 471},
  author={Your Name},
  year={2025},
  howpublished={\url{https://huggingface.co/ea-dev-final}}
}
```