---
license: apache-2.0
datasets:
- mayiwen/PaperAudit_Dataset
language:
- en
base_model:
- Qwen/Qwen3-14B
---
# PaperAudit Qwen3 14B (SFT + RL)

## Model Overview

PaperAudit_Qwen3_14B_sft_rl is a large-scale high-performance model specifically trained for academic paper error detection and automated review tasks. This model is based on Qwen3 14B and has been optimized through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), providing the strongest performance for complex academic paper analysis tasks.

## Model Information

- **Base Model**: Qwen3 14B
- **Model Parameters**: ~14 billion parameters
- **Training Method**: Supervised Fine-Tuning (SFT) + Reinforcement Learning (RLHF)
- **Model Architecture**: Qwen3ForCausalLM
- **Context Length**: 40,960 tokens
- **Data Type**: bfloat16

## Model Features

- **Top-Tier Performance**: 14B parameter scale, providing optimal performance when handling complex paper analysis tasks
- **Specialized Optimization**: Specifically optimized for academic paper error detection and review tasks
- **Reinforcement Learning**: Aligned with human preferences through RLHF to improve review quality and error detection accuracy
- **Long Context Support**: Supports 40K tokens context length, suitable for processing complete academic papers
- **Deep Understanding**: Capable of understanding complex academic concepts and writing norms

## Training Data

This model is trained on [PaperAudit_Dataset](https://huggingface.co/datasets/mayiwen/PaperAudit_Dataset). The dataset includes:

- Academic papers downloaded from OpenReview
- Structured paper content (processed via LlamaParse and LLM)
- Synthetic error data for training error detection models
- Human review feedback data

For more details about the dataset, please visit: https://huggingface.co/datasets/mayiwen/PaperAudit_Dataset

## Usage

### Install Dependencies

```bash
pip install transformers torch accelerate
```

### Load Model

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_path = "./qwen3_14b_sft_rl"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
```

### Inference Example

```python
# Prepare input (paper error detection task)
prompt = """Please detect errors in the following academic paper paragraph:

[Paper content...]

Please identify errors and provide correction suggestions."""

# Encode input
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate response
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True,
        pad_token_id=tokenizer.pad_token_id
    )

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

## Application Scenarios

- Academic paper error detection (highest accuracy)
- Automated paper review (professional-grade)
- Academic writing quality assessment (deep analysis)
- Paper content analysis and feedback generation (detailed feedback)
- Academic review assistant tools (expert-level advice)
- Complex academic concept understanding and analysis

## Model Architecture Details

- **Hidden Size**: 5120
- **Intermediate Size**: 17408
- **Number of Attention Heads**: 40
- **Number of Key-Value Heads**: 8 (Grouped Query Attention)
- **Number of Hidden Layers**: 40
- **Vocabulary Size**: 151,936

## Performance Advantages

Compared to the 8B and 3B models, the 14B model excels in the following aspects:

- **Higher Accuracy**: Capable of identifying more subtle and complex academic errors
- **Deeper Analysis**: Provides more detailed and professional review comments
- **Better Understanding**: Deeply understands academic writing norms, research methods, and theoretical frameworks
- **Stronger Reasoning**: Capable of complex logical reasoning and critical analysis
- **More Comprehensive Feedback**: Not only identifies errors but also provides constructive improvement suggestions

## Suitable Scenarios

Most suitable for:
- Paper review for high-quality academic journals
- Quality assessment of theses and dissertations
- Paper review for academic conferences
- Scenarios requiring extremely high review quality

## Notes

- This model is specifically optimized for academic paper review tasks and may require further fine-tuning for other domains
- It is recommended to use bfloat16 precision to save memory and improve inference speed
- For long document processing, appropriate context window management strategies are recommended
- Requires at least 28GB GPU memory for inference, multi-GPU inference or quantization techniques are recommended
- Inference speed is relatively slow, suitable for scenarios with high quality requirements and relatively lower speed requirements

## System Requirements

- **Recommended Memory**: At least 28GB GPU memory
- **Recommended Configuration**: A100 40GB or higher configuration
- **Quantization Options**: 8-bit or 4-bit quantization can be used to reduce memory requirements

## Related Resources

- **Training Dataset**: [PaperAudit_Dataset](https://huggingface.co/datasets/mayiwen/PaperAudit_Dataset)
- **PaperAudit Project**: For more details, please refer to the PaperAudit project documentation

## License

Please refer to the license terms of the base model Qwen3.