surindersinghssj/surt-whisper-small-pilot-phase1

Surt — Whisper Small Pilot Phase 1 (Text LM Priming)

LoRA adapter for openai/whisper-small, trained on Gurbani text-only data.

Training Details

• Phase: 1 — Text LM Priming (decoder self-attention only)
• LoRA targets: decoder self-attn Q, K, V projections (rank 16)
• Data: Gurbani text corpus from surindersinghssj/gurbani-asr-text
• Input: Dummy mel spectrograms (text-only training, no audio)
• Steps: 1,000 | Batch size: 32 | LR: 1e-4 (cosine decay)

Results

Metric	Value
Final train loss	2.42
Eval loss	1.23
Train loss (avg)	2.894

Purpose

Phase 1 primes the decoder to produce Gurmukhi text sequences before introducing audio. This checkpoint serves as the starting point for Phase 2 (audio fine-tuning).

W&B Run

surt-pilot/qnssfjr9