phuy1125/xlmr-large-fakenews-lora

XLM-RoBERTa-Large LoRA adapter (Stage-3, seed 42) — Cross-lingual Fake-News

LoRA adapter trained on the encoder backbone of an XLM-RoBERTa-Large teacher for cross-lingual (EN → VI) fake-news / claim verification (Stage-3 "mixed" run).

• Base model: xlm-roberta-large
• PEFT type: LoRA — r=8, lora_alpha=16, lora_dropout=0.2
• Target modules: query, value (task_type FEATURE_EXTRACTION)
• Teacher vi_test F1 (after merge): 0.8858
• Labels: 0 = fake, 1 = real

⚠️ Important

This adapter contains only the backbone LoRA delta. The classification head (a 2-layer MLP) was frozen during Stage-3 and is NOT included here, so the adapter alone does not perform classification end-to-end. It is published for reproducibility/archival. For ready-to-use classification, use the distilled student models which bundle the full head:

• phuy1125/xlmr-base-fakenews-kd
• phuy1125/mminilm-l6-fakenews-kd
• phuy1125/mminilm-l12-fakenews-kd

Usage (attach adapter to the backbone)

from transformers import AutoModel
from peft import PeftModel

backbone = AutoModel.from_pretrained("xlm-roberta-large")
backbone = PeftModel.from_pretrained(backbone, "phuy1125/xlmr-large-fakenews-lora")
backbone = backbone.merge_and_unload()  # fold LoRA into base weights