dehanns/whisper-small-16-lora-cs-42

Independent code-switch LoRA adapter

This is a standalone rank-16 code-switching LoRA adapter for openai/whisper-small. It was trained directly from the frozen base model using the new rich-codeswitched/train.csv split. The Sinhala rank-32 adapter was not loaded, changed, or retrained.

The best epoch was selected only by validation WER. The untouched test split was evaluated once after selection.

Rich code-switched test

• Samples: 62
• WER: 25.6550%
• CER: 11.5635%

Load this adapter on top of openai/whisper-small with PEFT. For the independent-weighted experiment, combine this adapter with the separate rank-32 Sinhala adapter only at inference time using validation-selected alpha/beta weights.