dehanns/whisper-small-sequential-cs-adapter

Sequential Sinhala + code-switch LoRA adapter

This repository contains the code-switch residual adapter from the sequential strategy. It was trained while the Sinhala adapter remained active and frozen.

It is not a standalone LoRA. Inference requires:

• base model: openai/whisper-small
• Sinhala adapter: dehanns/whisper-small-lora-sinhala-32 at revision epoch-5
• this CS residual adapter: dehanns/whisper-small-sequential-cs-adapter
• both adapters active with unit weight (SI=1.0, CS=1.0)

Held-out rich code-switched test result

• Samples: 62
• WER: 17.6856%
• CER: 7.2448%
• MER: 17.4006%

The test split was not used for training, model selection, or epoch selection. See sequential_dependency.json, sequential_train_report.json, and rich_codeswitched_test_metrics.json for the exact dependency and metrics.