XLMRShield
XLM-RoBERTa-base dual-head classifier for content moderation and prompt injection detection.
Fine-tuned with 40:60 resampling + per-class cap (30K) + WeightedRandomSampler.
Test Results
| Metric |
Score |
| Safety Accuracy |
0.8561 |
| Safety F1 |
0.5321 |
| Category Macro F1 |
0.6116 |
| Category Micro F1 |
0.6502 |
Per-Class F1
| Category |
F1 |
| benign |
0.5339 |
| child_sexual_exploitation |
0.4491 |
| hate_and_harassment |
0.6302 |
| indiscriminate_weapons |
0.5208 |
| misinformation_and_specialized_advice |
0.4748 |
| non_violent_crimes |
0.7812 |
| pi_and_jailbreak |
0.6587 |
| privacy |
0.6972 |
| sexual_content |
0.6889 |
| suicide_and_self_harm |
0.5734 |
| violent_crimes |
0.7195 |
Architecture
- Backbone: xlm-roberta-base (278M params)
- Head 1: Safety binary (sigmoid)
- Head 2: 11-class multi-label (sigmoid)
- Tokenizer: xlm-roberta-base
Training
- Dataset: satyamsaf3ai/merged_content_moderation_and_prompt_injection_new
- Balancing: 40:60 resampling + per-class cap 30K + WeightedRandomSampler
- BF16 mixed precision, AdamW + cosine LR, 4 epochs
- max_length=128, batch_size=32, lr=2e-05
