---
license: cc-by-nc-4.0
---
# XLMR-SocMed-UkrWarNarrative-5C_v1

## Model Description

**XLMR-SocMed-UkrWarNarrative-5C_v1** is a fine-tuned XLM-RoBERTa model designed for multi-class classification of social media narratives related to the Ukrainian conflict. The model identifies and classifies disinformation and propaganda patterns in social media discourse across 5 distinct categories.

## Model Details

- **Model Type:** XLM-RoBERTa (multilingual transformer)
- **Language(s):** Multilingual (optimized for social media content)
- **License:** cc-by-nc-4.0
- **Finetuned from:** xlm-roberta-large
- **Domain:** Social Media, Conflict Analysis, Narrative Classification
- **Classes:** 5 categories
- **Version:** 1.0


### Primary Use Cases
- Academic research on conflict-related social media discourse
- Content moderation and disinformation detection
- Media literacy and narrative analysis studies
- Social media monitoring for Ukrainian conflict-related content

## Performance

### Test Set Results
- **Accuracy:** 76.16%
- **Macro F1:** 77.51%
- **Macro Precision:** 75.06%
- **Macro Recall:** 81.28%

### Per-Class Performance

| Label | Category | Precision | Recall | F1-Score | Support |
|-------|----------|-----------|---------|----------|---------|
| 1601  | Anti-Western Narratives | 0.76 | 0.75 | 0.75 | 174 |
| 1602  | Economic Fallout/Domestic Welfare Neglected | 0.69 | 0.88 | 0.77 | 50 |
| 1603  | Illegitimate and Corrupt Ukrainian Leadership | 0.72 | 0.83 | 0.77 | 76 |
| 1604  | Ukraine/Nazi Allegations | 0.77 | 0.95 | 0.85 | 64 |
| 1699  | None (Legitimate Discourse) | 0.81 | 0.65 | 0.72 | 194 |

## Training Details

### Training Data
- **Source:** Social media platforms (Twitter/X)

### Training Procedure
- **Epochs:** 4
- **Final Training Loss:** 0.3397
- **Validation Accuracy:** 78.63%
- **Validation F1 (Macro):** 79.93%

## Label Descriptions

### 1601 - Anti-Western Narratives
Consolidates NATO expansion criticism, Western interference allegations, historical grievances, propaganda claims, and Russian defense justifications. This unified framework demonstrates substantial empirical presence across 13 clusters (~583,000 tweets) with strong thematic coherence and overlapping linguistic patterns.

### 1602 - Economic Fallout/Domestic Welfare Neglected
Critical narratives focusing on resource allocation to Ukraine at the expense of domestic priorities. Shows strong empirical presence across 2 clusters (~90,000 tweets), emphasizing economic concerns and domestic welfare issues.

### 1603 - Illegitimate and Corrupt Ukrainian Leadership  
Narratives aimed at delegitimizing Ukrainian government through corruption allegations and personal attacks. Shows concentrated empirical evidence despite limited keyword matches, focusing on undermining leadership credibility.

### 1604 - Ukraine/Nazi Allegations
The most empirically robust category with presence across 8 clusters (~157,000 tweets). Frames Ukraine as ideologically aligned with Nazi principles to justify Russian intervention. Shows highest recall (0.95) indicating strong pattern recognition.

### 1699 - None (Legitimate Discourse)
Captures legitimate discourse, factual reporting, and content outside the scope of the above narrative categories. Represents balanced, non-propagandistic content related to the Ukrainian conflict.

## Bias and Limitations

### Known Limitations
- Model performance varies across categories (F1: 0.72-0.85)
- Lower recall for "None" category (0.65) may misclassify legitimate content
- Training data reflects social media biases and temporal patterns

## Ethical Considerations

### Responsible Use
- Requires human oversight for content moderation decisions
- Should not be used as sole authority for censorship decisions  
- Results should be interpreted within broader context of media analysis
- Regular monitoring for performance drift and bias amplification

### Potential Risks
- False positive classifications may affect legitimate speech
- Model may reflect historical biases present in training data
- Adversarial attacks could exploit classification boundaries

## Technical Specifications

### Model Architecture
- Base Model: xlm-roberta-large
- Fine-tuning: Classification head with 5 output classes
- Input: Social media text (max length as per XLM-RoBERTa specifications)
- Output: Class probabilities for 5 narrative categories

### Hardware Requirements
- GPU recommended for inference at scale
- Standard transformer model memory requirements
- Optimized for batch processing of social media content

## Citation

```bibtex
@misc{xlmr-socmed-ukrwarnarrative-5c-v1,
  title={XLMR-SocMed-UkrWarNarrative-5C_v1: A Multilingual Classifier for Ukrainian Conflict Narratives on Social Media},
  author={[Your Name/Institution]},
  year={2025},
  url={[Model URL]}
}
```

## Model Card Authors
[Orsolya Ring/poltextLAB]



---
*This model card follows the guidelines established by Mitchell et al. (2019) for transparent AI model documentation.*