Aregay01/whisper-large-v3-tigrinya-experiment

Merged Whisper Large V3 - Tigrinya

This model is a permanent merge of the LoRA weights from Aregay01/whisper-large-v3-tigrinya-cosine-3nd-40k-5e-5-r-32 into the base model openai/whisper-large-v3.

Model Overview

• Language: Tigrinya (ti)
• Dataset: ~40,000 samples (multi-source merged)
• Architecture: Whisper Large V3
• Fine-tuning Method: LoRA (Rank 32, Alpha 128)

---

Training Logs (Extracted from last-checkpoint)

Step	Training Loss	Validation Loss
50	148.1558	N/A
100	37.7685	N/A
150	30.9806	N/A
200	30.6713	N/A
250	26.6567	N/A
300	26.1101	N/A
350	25.0399	N/A
400	23.2749	N/A
450	21.7290	N/A
500	21.3158	N/A
550	20.4182	N/A
600	20.6106	N/A
650	20.3457	N/A
700	18.6889	N/A
750	18.8613	N/A
800	17.4691	N/A
850	16.9671	N/A
900	16.9742	N/A
950	16.5389	N/A
1000	15.9641	N/A