README.md

---
language: en
license: mit
tags:
- content-safety
- text-classification
- safety
- moderation
- deberta
datasets:
- jainsatyam26/guardrail-215k-splits
metrics:
- f1
- accuracy
widget:
- text: "How to make a bomb?"
  example_title: "Violent Content"
- text: "Hello, how are you?"
  example_title: "Safe Content"
---

# High-Accuracy Content Safety Classifier

This model is a fine-tuned DeBERTa classifier for content safety, achieving high performance on safety classification tasks.

## Model Details

- **Base Model**: microsoft/deberta-v3-large
- **Training Dataset**: jainsatyam26/guardrail-215k-splits
- **Categories**: 10 safety categories
- **Training Time**: Auto-deployed during training
- **Last Updated**: 2026-04-29 06:21:01 UTC


## Performance

| Metric | Value |
|--------|-------|
| F1 Score | N/A |
| Accuracy | N/A |
| Unsafe F1 | N/A |


## Categories

- `benign`
- `jailbreak`
- `S1 Violent Crimes`
- `S2 Non-Violent Crimes`
- `S4 Child Sexual Exploitation`
- `S7 Privacy`
- `S10 Hate`
- `S11 Self-Harm`
- `S12 Sexual Content`
- `S14 Code Abuse`

## Usage

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

tokenizer = AutoTokenizer.from_pretrained("jainsatyam26/bertclassfier")
model = AutoModelForSequenceClassification.from_pretrained("jainsatyam26/bertclassfier")

def predict(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=-1)
        predicted_id = torch.argmax(probs, dim=-1).item()
    
    labels = ['benign', 'jailbreak', 'S1 Violent Crimes', 'S2 Non-Violent Crimes', 'S4 Child Sexual Exploitation', 'S7 Privacy', 'S10 Hate', 'S11 Self-Harm', 'S12 Sexual Content', 'S14 Code Abuse']
    return {
        "prediction": labels[predicted_id],
        "confidence": probs[0][predicted_id].item(),
        "all_scores": {labels[i]: probs[0][i].item() for i in range(len(labels))}
    }

# Example
result = predict("How to make a bomb?")
print(result)
```

## Training Configuration

This model was trained with the following configuration:

```json
{
  "model_name": "microsoft/deberta-v3-large",
  "dataset_name": "jainsatyam26/guardrail-215k-splits",
  "max_length": 512,
  "epochs": 4,
  "batch_size": 8,
  "grad_accum": 4,
  "learning_rate": 1e-05,
  "weight_decay": 0.01,
  "warmup_ratio": 0.1,
  "use_llrd": true,
  "llrd_alpha": 0.9,
  "use_multisample_dropout": true,
  "num_dropout_samples": 5,
  "dropout_rate": 0.3,
  "use_label_smoothing": true,
  "label_smoothing": 0.1,
  "use_focal_loss": true,
  "focal_alpha": 0.7,
  "focal_gamma": 2.0,
  "use_hard_negative": true,
  "hard_negative_ratio": 0.3,
  "num_folds": 3,
  "optimize_thresholds": true,
  "output_dir": "./guardrail_model",
  "checkpoint_steps": 500,
  "logging_steps": 50,
  "eval_steps": 500,
  "hf_repo_id": "jainsatyam26/bertclassfier",
  "hf_token": "***REDACTED***",
  "deploy_every_minutes": 30,
  "deploy_every_steps": 400,
  "auto_deploy": true,
  "private_repo": false,
  "auto_resume": true,
  "resume_from_hf": true,
  "use_wandb": true,
  "wandb_project": "safety-classifier",
  "fp16": false,
  "bf16": true,
  "dataloader_num_workers": 4,
  "seed": 42
}
```

## Automatic Deployment

This model is automatically deployed every 30 minutes during training with:
- ✅ Automatic checkpoint recovery
- ✅ Real-time performance monitoring  
- ✅ Progressive model updates
- ✅ Training state persistence

---

*Generated automatically during training - 2026-04-29 06:21:01*