metadata
language:
- fr
license: mit
base_model: MEscriva/gilbert-fr-source
tags:
- whisper
- asr
- french
- mathematics
- education
- lora
- fine-tuned
pipeline_tag: automatic-speech-recognition
datasets:
- Lexia-Labs/french-math-asr-benchmark
Gilbert-FR-Math v1
Whisper Large V3 fine-tune pour la transcription de cours de mathematiques en francais.
Entrainement
- Base: MEscriva/gilbert-fr-source
- Methode: LoRA r=32 alpha=64 + 4-bit NF4
- Dataset: Lexia-Labs/french-math-asr-benchmark
- Epochs: 3, batch 4x4
- GPU: NVIDIA T4 16GB