voxtral-finetune-20250913_164641

Fine-tuned Voxtral ASR model

Usage

import torch
from transformers import AutoProcessor, AutoModelForSeq2SeqLM
import soundfile as sf

processor = AutoProcessor.from_pretrained("Tonic/voxtral-finetune-20250913_164641")
model = AutoModelForSeq2SeqLM.from_pretrained(
    "Tonic/voxtral-finetune-20250913_164641",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

audio, sr = sf.read("sample.wav")
inputs = processor(audio, sampling_rate=sr, return_tensors="pt")
with torch.no_grad():
    generated_ids = model.generate(**inputs, max_new_tokens=256)
text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(text)

Training Configuration

• Base model: mistralai/Voxtral-Mini-3B-2507
• Config: Custom Configuration
• Trainer: SFTTrainer

Training Parameters

• Batch size: 2
• Grad accumulation: 4
• Learning rate: 5e-05
• Max epochs: 3
• Sequence length: 2048

Hardware

• GPU: NVIDIA RTX 4000 Ada Generation

Notes

• This repository contains a fine-tuned Voxtral ASR model.