maryamas222/whisper-large-v3-egyptian-lora

basemodel: openai/whisper-large-v3 libraryname: peft

language:

• —ar

Training Details

• —Learning rate: 1e-5
• —Batch size: 4 (with gradient accumulation steps of 2)
• —Epochs: 20 (early stopping with patience=5)
• —Best checkpoint: step 864
• —Warmup steps: 20
• —Precision: bfloat16
• —Hardware: NVIDIA A100 80GB

Training Data

MightyStudent/Egyptian-ASR-MGB-3

Results

WER (MGB-3 test split) | 31.06% |

Evaluation was performed with Arabic text normalization applied (diacritics removed, alef variants normalized, punctuation removed).

• —PEFT 0.7.1