Downloads · 30 days
21
7% of all-time downloads
xlm-roberta-factchecker
Ashg2099/xlm-roberta-factchecker
xlm-roberta-factchecker is a text classification model from Ashg2099. Use it when you need a label for a piece of text. The card lists the license as mit.
Fine-tuned xlm-roberta-base for multilingual fact verification using FEVER + BGE retrieval.
All-time downloads
290
Public
Parameters
278M
3.4 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors1.1 GB · 98%
At a glance
- Task
- Text Classification
- License
- mit
- Model type
- xlm-roberta
- Access
- Public
- Created
- Jun 4, 2026
- Updated
- Jun 19, 2026
- SHA
- 19176706
Datasets
Model card
From the Hugging Face model README
XLM-RoBERTa Fact Checker (v2 — Retrieval Augmented)
Fine-tuned xlm-roberta-base for multilingual fact verification using FEVER + BGE retrieval.
What's New in v2
Previous version (claim only) achieved F1 0.655 but misfired on real-world claims. v2 feeds claim + retrieved evidence to the model, fixing the core limitation.
Results
| Metric | v1 (claim only) | v2 (claim + evidence) |
|---|---|---|
| F1 macro | 0.655 | 0.813 |
| Accuracy | 0.63 | 0.79 |
Real World Performance
| Claim | v1 | v2 |
|---|---|---|
| Earth is 3rd planet from Sun | FALSE ❌ | TRUE ✅ |
| Eiffel Tower in Berlin | TRUE ❌ | FALSE ✅ |
| WW2 ended in 1945 | FALSE ❌ | TRUE ✅ |
| Obama was 44th President | TRUE ✅ | TRUE ✅ |
Architecture
- Base model: xlm-roberta-base
- Retriever: BAAI/bge-base-en-v1.5
- Index: FAISS (56,946 FEVER evidence passages)
- Input format:
[claim] </s></s> [retrieved evidence] - Training: 1 fine-tuning epoch on claim+evidence pairs, lr=5e-6
Labels
| ID | Label | Meaning |
|---|---|---|
| 0 | TRUE | Claim supported by evidence |
| 1 | FALSE | Claim refuted by evidence |
| 2 | UNCERTAIN | Insufficient evidence |
Known Limitations
This model requires evidence at inference time. Use BGE retrieval (BAAI/bge-base-en-v1.5) to retrieve relevant passages before classification. Standalone claim input without evidence will underperform.
Part of Sift
This model is part of Sift — an open-source 5-agent fact-checking pipeline.
- Task
- Text Classification
- Type
- xlm-roberta
- License
- mit
- Languages
- multilingual
- Created
- Jun 4, 2026
- Updated
- Jun 19, 2026