Modelpublic
h-gajdov/whisper_full_finetune_all_dialects_lr1e-5
0likes137downloads
Model Card
Results
| Model | Config | WER ↓ | CER ↓ | Word Acc ↑ | Char Acc ↑ | BLEU ↑ | ROUGE-1 ↑ | ROUGE-2 ↑ | ROUGE-L ↑ | RTF ↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| Base Whisper Large V3 Turbo | None | 45.53% | 26.62% | 61.36% | 79.79% | 42.68 | 61.79 | 45.23 | 61.43 | 0.00636 |
| whisper_full_finetune_all_dialects_lr1e-5 | Full FT, lr=1e-5 | 16.68% | 9.09% | 87.90% | 94.32% | 77.79 | 89.49 | 82.44 | 89.34 | 0.00486 |
| whisper_lora_all_dialects_r32_qk_lr1e-3 | r=32, Q/K, lr=1e-3 | 16.56% | 9.14% | 87.62% | 94.19% | 77.60 | 89.11 | 81.83 | 88.96 | 0.00677 |
| whisper_lora_all_dialects_r32_qkvo_lr5e-4 | r=32, Q/K/V/O, lr=5e-4 | 18.78% | 10.10% | 85.73% | 93.42% | 74.37 | 87.28 | 79.00 | 87.11 | 0.00662 |
| whisper_lora_all_dialects_r64_qkvo_lr5e-4 | r=64, Q/K/V/O, lr=5e-4 | 15.27% | 8.26% | 87.98% | 94.44% | 78.15 | 89.22 | 81.95 | 89.08 | 0.00649 |
Eval Diagram
[image]
Train diagram
[image]