ParaRater: Enhancing Cross-Lingual Transfer in LLMs with Meta-Learning

ParaRater is a data selection method that enhances cross-lingual transfer by selecting the most valuable parallel pairs, forming high-impact parallel corpora with two meta-learned raters.

ParaTool

This is the meta-learning toolkit repository for training raters.

Data Format

Use Parquet files with the following columns:

• text (required): string
• category (optional): string label used only for logging per-category losses

Train

accelerate launch --num_processes 7 pararater.py \
  --trainset_path /path_to/train \
  --valset_path   /path_to/val