TR-MTEB retrieval, mini held-out sets + training remainder
Small, fast Turkish retrieval benchmarks cut from the TR-MTEB retrieval tasks, built for embeddingmagibu2. Every config has one split per task.
queries,corpus: held-out evaluation. 200 random test queries per task (seed 0); the corpus holds all their relevant documents, the top-20 BM25 documents of each query (hard distractors) and random documents up to 10,000 (the whole corpus when it is smaller). In ArguAna and Quora the held-out queries are removed from the corpus unless relevant. Never train on these.train: every other qrel of the task (train, dev and the remaining test queries) with score > 0, minus every pair whose query or positive text is a held-out query or a held-out positive. Safe to train on next to the held-out sets.
Scores on these mini sets are not comparable with full TR-MTEB numbers (smaller corpus); they are for comparing models.
| Split | Source | Held-out queries | Corpus (mini / full) | Train pairs | Qrels dropped for leakage |
|---|---|---|---|---|---|
msmarco |
trmteb/msmarco-tr | 200 | 10,000 / 718,209 | 316,290 | 242 |
fiqa |
trmteb/fiqa-tr | 200 | 10,000 / 57,638 | 16,596 | 514 |
nfcorpus |
trmteb/nfcorpus-tr | 200 | 3,633 / 3,633 | 13,490 | 120,804 |
scifact |
trmteb/scifact-tr | 200 | 5,183 / 5,183 | 798 | 460 |
arguana |
trmteb/arguana-tr | 200 | 8,486 / 8,674 | 1,177 | 224 |
scidocs |
trmteb/scidocs-tr | 200 | 10,000 / 25,657 | 3,402 | 1,526 |
quora |
trmteb/quora-tr | 200 | 10,000 / 522,931 | 22,961 | 340 |
cqadupstack_gaming |
trmteb/cqadupstack-gaming-tr | 200 | 10,000 / 45,301 | 1,993 | 270 |
squad |
trmteb/squad-tr | 200 | 1,204 / 1,204 | 6,878 | 1,413 |
tquad |
trmteb/tquad | 200 | 275 / 275 | 292 | 600 |
Columns: queries: id, text, pos_ids, pos_scores · corpus: id, title, text · train: query, title, text (positive document), score.
Script: make_eval_set.py in the embeddingmagibu2 project. Licenses follow the source datasets.
- Downloads last month
- 66