| --- |
| configs: |
| - config_name: en-tl |
| data_files: |
| - split: train |
| path: paracore_en-tl/*.parquet |
| - config_name: en-zh |
| data_files: |
| - split: train |
| path: paracore_en-zh/*.parquet |
| - config_name: en-th |
| data_files: |
| - split: train |
| path: paracore_en-th/*.parquet |
| - config_name: en-ru |
| data_files: |
| - split: train |
| path: paracore_en-ru/*.parquet |
| - config_name: en-ja |
| data_files: |
| - split: train |
| path: paracore_en-ja/*.parquet |
| - config_name: en-fr |
| data_files: |
| - split: train |
| path: paracore_en-fr/*.parquet |
| - config_name: en-de |
| data_files: |
| - split: train |
| path: paracore_en-de/*.parquet |
| - config_name: en-ar |
| data_files: |
| - split: train |
| path: paracore_en-ar/*.parquet |
| --- |
| |
| # 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. |
|
|
| # ParaCore |
|
|
| This is the parallel data filtered by ParaRater. English data is sampled from Common Crawl. The translation is conducted using `Qwen3-8B`. Each language contains about 1 B tokens. |
