---
language:
- en
tags:
- fact-checking
- misinformation-detection
- bert
- modernbert
datasets:
- FELM
- FEVER
- HaluEval
- LIAR
metrics:
- accuracy
- f1
---

# ModernBERT Fact-Checking Model

## Model Description

This is a fine-tuned ModernBERT model for binary fact-checking classification, trained on consolidated datasets from multiple authoritative sources. The model determines whether a given claim is likely to be true (label 1) or false (label 0).

**Base Model:** [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base)

## Intended Uses

### Primary Use
- Automated fact-checking systems
- Misinformation detection pipelines
- Content moderation tools

### Out-of-Scope Uses
- Multilingual fact-checking (English only)
- Medical/legal claim verification
- Highly domain-specific claims

### How to use

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("your-username/modernbert-factchecking")
model = AutoModelForSequenceClassification.from_pretrained("your-username/modernbert-factchecking")

inputs = tokenizer("Your claim to verify here", return_tensors="pt", truncation=True, max_length=512)
outputs = model(**inputs)
predictions = torch.softmax(outputs.logits, dim=-1)
```

## Training Data

The model was trained on a combination of four datasets:

| Dataset | Samples | Domain |
|---------|---------|--------|
| FELM | 34,000 | General claims |
| FEVER | 145,000 | Wikipedia-based claims |
| HaluEval | 12,000 | QA hallucination detection |
| LIAR | 12,800 | Political claims |

**Total training samples:** ~203,800

## Training Procedure

### Hyperparameters
- Learning Rate: 5e-5
- Batch Size: 32
- Epochs: 1
- Max Sequence Length: 512 tokens
- Optimizer: adamw_torch_fused

### Preprocessing
All datasets were converted to a standardized format:
```python
{
  "text": "full claim text",
  "label": 0.0 or 1.0,
  "source": "dataset_name"
}