README.md

---
license: mit
language:
- en
base_model:
- Qwen/Qwen3-Embedding-0.6B
pipeline_tag: text-classification
---

# Argus

**AI-Generated Text Detection Classifier**

Argus is a binary text classifier that detects whether text was written by a human or generated by AI. It is fine-tuned from [Qwen/Qwen3-Embedding-0.6B](https://huggingface.co/Qwen/Qwen3-Embedding-0.6B) using a classification head with last-token pooling.

## Features

- **High Accuracy**: Achieves near-perfect classification on held-out test data
- **Long Context**: Supports sequences up to 4,096 tokens with automatic chunking for longer texts
- **Fast Inference**: Optimized with Flash Attention 2 and bfloat16 precision on CUDA
- **Batch Processing**: Parallel tokenization and batched inference for high throughput

## Installation

```bash
pip install torch transformers
```

## Minimal Inference Example

```python
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer, AutoConfig

class Qwen3ForSequenceClassification(nn.Module):
    """Qwen3-Embedding with classification head using last-token pooling."""
    
    def __init__(self, model_name="Qwen/Qwen3-Embedding-0.6B", num_labels=2):
        super().__init__()
        self.encoder = AutoModel.from_pretrained(
            model_name,
            torch_dtype=torch.bfloat16,
            trust_remote_code=True,
        )
        hidden_size = AutoConfig.from_pretrained(model_name).hidden_size
        self.classifier = nn.Linear(hidden_size, num_labels)
    
    def forward(self, input_ids, attention_mask):
        outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        # Last-token pooling (not CLS token)
        pooled = outputs.last_hidden_state[:, -1]
        return self.classifier(pooled)


# Load model and weights
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Qwen3ForSequenceClassification()
model.load_state_dict(torch.load("weights/model.pt", map_location=device, weights_only=True))
model.to(device).eval()

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("weights/", padding_side="left", trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Inference
text = "Your text to classify here."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=4096)
inputs = {k: v.to(device) for k, v in inputs.items()}

with torch.no_grad():
    logits = model(inputs["input_ids"], inputs["attention_mask"])
    probs = F.softmax(logits.float(), dim=-1)[0]
    
label = "ai" if probs[1] > probs[0] else "human"
confidence = probs[1].item() if label == "ai" else probs[0].item()

print(f"Label: {label}, Confidence: {confidence:.2%}")
# Example output: Label: human, Confidence: 94.32%
```

## Model Architecture

| Component | Details |
|-----------|---------|
| Base Model | Qwen/Qwen3-Embedding-0.6B |
| Hidden Size | 1024 |
| Parameters | ~600M |
| Pooling | Last-token (not CLS) |
| Classification Head | Linear (1024 → 2) |
| Precision | bfloat16 (CUDA) / float32 (CPU) |

## Performance

| Metric | Score |
|--------|-------|
| Accuracy | 98.86% |

*Note: These metrics are from the validation set during training. Real-world performance may vary depending on the domain and AI models used to generate text.*

## Training Data

Trained on a combination of:
- [RAID](https://huggingface.co/datasets/liamdugan/raid) - Multi-domain, multi-model dataset
- [HC3](https://huggingface.co/datasets/Hello-SimpleAI/HC3) - Human vs ChatGPT responses
- [AI-human-text](https://huggingface.co/datasets/andythetechnerd03/AI-human-text)
- [AI Text Detection Pile](https://huggingface.co/datasets/artem9k/ai-text-detection-pile)

## License

MIT License

## Citation

If you use Argus in your research, please cite:

```bibtex
@software{,
  title={Argus: AI-Generated Text Detection Classifier},
  author={Xi Nai Lai},
  year={2026},
  url={https://huggingface.co/johnbean393/argus/}
}
```