← All workNLP / Agents
TruthShield — Agentic Fake News Detection
Fine-tuned RoBERTa with a GPT-2 perplexity fusion head, paired with an LLM agent that checks claims against live news — explainable via attention rollout.
- Stack
- RoBERTa · LLM Agent · FastAPI · Modal
- Key figure
- 0.9956 — Validation F1
DG Labs
TruthShield
0.9956Validation F1
- RoBERTa + GPT-2 perplexity fusion head
- Validation F1 0.9956
- LLM agent checks claims against live news
RoBERTa · LLM Agent · FastAPI · Modal
01
The problem
A classifier can say a story looks fake; it cannot say why, or check it against what is happening right now.
02
The approach
Fine-tuned RoBERTa with a GPT-2 perplexity fusion head, and paired it with an LLM agent that checks claims against live news. Attention rollout shows which words drove each verdict.
03
The outcome
Validation F1 of 0.9956, deployed live on Modal.
Of note
- ✦ RoBERTa + GPT-2 perplexity fusion head
- ✦ Validation F1 0.9956
- ✦ LLM agent checks claims against live news
- ✦ Explainable via attention rollout