IonGrozea/whisper-small_ro-80mel

Fine-tuned from openai/whisper-small on the ALL_RO_80MEL Romanian dataset.

Training setup

• Base model: openai/whisper-small
• Dataset: ALL_RO_80MEL (train / validation / test)
• Objective: ASR (Romanian transcription)
• Trainer: Hugging Face Seq2SeqTrainer
• Checkpoint saved: best WER on validation (final_best/)

Training results (per evaluation)

epoch	step	train_loss	eval_loss	eval_wer	eval_cer
1.0000	6475		0.1062	2.8610	2.9901
2.0000	12950		0.0852	3.8649	3.7047
3.0000	19425		0.0792	4.4320	4.2204
4.0000	25900		0.0857	4.3639	4.3786
4.0000	25900		0.1062	2.6983	2.9504

Metrics columns:

• train_loss – training loss logged near this eval step
• eval_loss – validation loss from trainer.evaluate()
• eval_wer – validation WER on a subset (lower is better)
• eval_cer – validation CER on a subset (lower is better)

The raw HF Trainer logs are also stored in:

• training_log.jsonl
• trainer_state.json
• eval_results.json
• data_results.json