| configs: | |
| - config_name: corpus | |
| data_files: | |
| - split: train | |
| path: data/corpus/*.parquet | |
| - config_name: queries | |
| data_files: | |
| - split: train | |
| path: data/queries/*.parquet | |
| - config_name: qrels | |
| data_files: | |
| - split: train | |
| path: data/qrels/*.parquet | |
| # MehdiAstaraki/progressive-code-switch | |
| **Progressive code-switching** retrieval-decay benchmark. Each base patent yields a cumulative ladder of documents (`base__r0` clean → `base__rN`), where each step swaps one more chemistry term into another language / spelling / ChEBI form. One fixed question per base (about the step-1 term) is reused for every depth; the `qrels` carry the `depth` so you can measure how retrieval decays as more terms are code-switched. | |
| Configs: `corpus` (ladder variant documents, with `depth` + `mode_added`), `queries` (one per base), `qrels` (query→variant, score 1, with `depth`). The retrieval haystack is this `corpus` PLUS a shared patent corpus passed at eval time. | |