Tohirju/tajik-asr-full-data

Tajik ASR fine-tune dataset (prepared)

Prepared training data for fine-tuning Peacockery/tajik-parakeet-tdt-110m on Tajik. Sources: muhtasham tajik-asr (LevDist-filtered clean subset) + a ~15% replay slice of Peacockery/tajik-asr-corpus-v3 (anti-forgetting). Books/code-switch data excluded.

• audio.tar (or audio_partNN.tar shards — extract ALL of them into the same dir: for t in audio*.tar; do tar -xf "$t"; done) — yields wav/ (train+val) and fleurs_wav/ (held-out eval)
• train_rel.jsonl / val_rel.jsonl — NeMo manifests (relative paths; extract tar in same dir)
• fleurs_tg_test_rel.jsonl — held-out FLEURS-tg eval (NEVER train on it)
• train: 231175 clips | val: 7149 | fleurs eval: 600

Public but gated (manual approval). Derived from external ASR sources; attribution to muhtasham and Peacockery.