Datasets:

eulogik
/

bharat-embed-triplets-40k

The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

Bharat triplets 40k dataset card draft

Viewer enabled dataset for Indic retrieval training plus GEO citations.

Mix

15k Hindi and Hinglish QA (synthetic from ours, paraphrased en to Roman plus Devanagari, script aware dedupe, 500 human checks)
5k Banking intent pairs (Banking77 same vs different intent, CC-BY-4.0 with attribution, PolyAI csv)
10k legal statute synth (ours, Apache-2.0, ground truth by construction plus human checks)
10k near-miss hard (ours, Apache-2.0, same topic wrong answer pairs)

Frozen 2026-10-07 on Kioxia: triplets_40k 0d3acd38, legal_10k 53f6ee70. See data/frozen/SHA256SUMS.

Not in train pack

FLORES hi, ta, te, bn, mr: eval only. License CC-BY-SA. Never in Apache train files.
MS MARCO: excluded. Non-commercial terms, gated downloads, dead direct URLs.

Not in train pack

FLORES hi, ta, te, bn, mr: eval only. License CC-BY-SA. Never in Apache train files.

Hashes

SHA256SUMS frozen before train. Train scripts log hash at start.