hindko-embeddings / eval /selection.json
junaid008's picture
Hindko Embeddings v1.0.0: multilingual-e5-large fine-tuned for Hindko (model, card, benchmarks)
4734758 verified
Raw History Blame Contribute Delete
1.66 kB
{
"rule": "pool = round-2 dev results + incumbents ft-e5-base (size 2) and ft-e5-small (size 1) with their round-1 dev scores; best = max overall; near = overall >= best - 0.003; selected = smallest size rank in near (ties by higher score); runner-up = best-scoring other model",
"dev_scores": {
"ft-e5-base": 0.6851297501603847,
"ft-e5-small": 0.6700276166417629,
"r2-bgem3": 0.7007568407810918,
"r2-e5b": 0.7079351835925609,
"r2-e5b-plain": 0.7066095544829916,
"r2-e5l": 0.7117165057564516,
"r2-e5s": 0.7038754488016046
},
"near": [
"r2-e5l"
],
"selected": "r2-e5l",
"runner_up": "r2-e5b",
"tie": 0.003,
"written_utc": "2026-10-01T17:48:17Z",
"gate": {
"pass": true,
"decided_utc": "2026-10-01T15:38:50Z",
"BookTitleRetrieval": 0.446437329167755,
"overall": 0.7079351835925609,
"BookClustering": 0.7339101300956627
},
"pairs_sha256": {
"rows.jsonl": "9ddde8f6ef2750166e4d895b8aba55ba394ba954961169ab0aec0f5d4a541faa",
"texts.jsonl": "f0c31d84f79e7c3cee0b45e708ec1bb1929f9e2d99d7fbdec66ecb47a1024a28"
},
"size_rank": {
"r2-e5s": 1,
"ft-e5-small": 1,
"r2-e5b": 2,
"r2-e5b-plain": 2,
"ft-e5-base": 2,
"r2-e5l": 3,
"r2-bgem3": 4
},
"notes": "dev_scores of ft-e5-base and ft-e5-small are their round-1 dev results (max_seq_length 256, float32); all other pool members were evaluated at max_seq_length 512 in float16. Test was run once, after this selection, on the selected model and the runner-up. gate = the pre-registered recipe gate evaluated on r2-e5b dev (BookTitleRetrieval >= 0.36, overall >= 0.685, BookClustering >= 0.68); pass -> the large runs used mined hard negatives and same-book blocks."
}