README.md

---
license: apache-2.0
language:
- en
tags:
- text-classification
- propaganda-detection
- multi-label
- modernbert
datasets:
- synapti/nci-propaganda-production
- synapti/nci-synthetic-articles
metrics:
- f1
- precision
- recall
pipeline_tag: text-classification
library_name: transformers
base_model: answerdotai/ModernBERT-base
---

# NCI Technique Classifier v2

**Multi-label propaganda technique classifier** trained on the NCI (Neural Counter-Intelligence) Protocol dataset.

## Model Description

This model detects **18 propaganda techniques** in text using a multi-label classification approach. It is designed to work as Stage 2 in a two-stage pipeline:

1. **Stage 1**: Binary detection (is there propaganda?) using `synapti/nci-binary-detector`
2. **Stage 2**: Technique classification (what techniques are used?) using this model

### Supported Techniques

| Technique | Description |
|-----------|-------------|
| `Loaded_Language` | Words/phrases with strong emotional implications |
| `Appeal_to_fear-prejudice` | Building support by exploiting fear |
| `Exaggeration,Minimisation` | Making something more/less important than it is |
| `Repetition` | Repeating the same message over and over |
| `Flag-Waving` | Playing on national/group identity |
| `Name_Calling,Labeling` | Attacking through labels rather than arguments |
| `Reductio_ad_hitlerum` | Persuading by comparing to disliked groups |
| `Black-and-White_Fallacy` | Presenting only two choices |
| `Causal_Oversimplification` | Assuming single cause for complex issue |
| `Whataboutism,Straw_Men,Red_Herring` | Deflection and misdirection |
| `Straw_Man` | Misrepresenting someone's argument |
| `Red_Herring` | Introducing irrelevant topics |
| `Doubt` | Questioning credibility without evidence |
| `Appeal_to_Authority` | Relying on authority rather than evidence |
| `Thought-terminating_Cliches` | Phrases that discourage critical thought |
| `Bandwagon` | Appeals to popularity |
| `Slogans` | Brief, memorable phrases |
| `Obfuscation,Intentional_Vagueness,Confusion` | Deliberately unclear language |

## Usage

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

# Load model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained(
    "synapti/nci-technique-classifier-v2",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "synapti/nci-technique-classifier-v2",
    trust_remote_code=True
)

# Prepare input
text = "Wake up, patriots! The radical elites are destroying our country!"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)

# Get predictions
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits)[0]

# Get technique labels
id2label = model.config.id2label
threshold = 0.3

# Print detected techniques
for idx, prob in enumerate(probs):
    if prob.item() >= threshold:
        technique = id2label[str(idx)]
        print(f"{technique}: {prob.item():.1%}")
```

### Two-Stage Pipeline Usage

```python
from nci.transformers.two_stage_pipeline import TwoStagePipeline

# Load two-stage pipeline
pipeline = TwoStagePipeline.from_pretrained(
    binary_model="synapti/nci-binary-detector",
    technique_model="synapti/nci-technique-classifier-v2",
)

# Analyze text
result = pipeline.analyze("Some text to analyze...")
print(f"Has propaganda: {result.has_propaganda}")
print(f"Confidence: {result.propaganda_confidence:.1%}")
print(f"Detected techniques: {result.detected_techniques}")
```

## Training Details

### Training Data

- **Primary**: [synapti/nci-propaganda-production](https://huggingface.co/datasets/synapti/nci-propaganda-production) (11,573 samples)
- **Augmentation**: [synapti/nci-synthetic-articles](https://huggingface.co/datasets/synapti/nci-synthetic-articles) (~5,485 synthetic article-length samples)
- **Total**: ~17,000 training samples

### Training Procedure

- **Base model**: [answerdotai/ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base)
- **Fine-tuning**: HuggingFace AutoTrain on A100 GPU
- **Epochs**: 3
- **Batch size**: 16
- **Learning rate**: 2e-5
- **Loss function**: Focal Loss (gamma=2) for class imbalance handling

### Performance Metrics

**Test Set Performance:**

| Metric | Score |
|--------|-------|
| **Micro F1** | 80.1% |
| **Macro F1** | 51.2% |

**Top Performing Techniques:**

| Technique | F1 Score |
|-----------|----------|
| Loaded_Language | 97.0% |
| Appeal_to_fear-prejudice | 89.7% |
| Name_Calling,Labeling | 81.8% |
| Exaggeration,Minimisation | 75.4% |

## Limitations

- Trained primarily on English text
- Performance varies by technique (common techniques perform better)
- Best used as Stage 2 after binary detection for efficient inference
- Requires `trust_remote_code=True` for ModernBERT architecture

## Citation

If you use this model, please cite:

```bibtex
@misc{nci-technique-classifier-v2,
  title={NCI Technique Classifier v2: Multi-label Propaganda Detection},
  author={Synapti},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/synapti/nci-technique-classifier-v2}
}
```

## License

Apache 2.0