cohere-darija-lora-10-7-26

Cohere ASR - Moroccan Darija (LoRA Adapter)

This model is a fine-tuned LoRA adapter for the CohereLabs/cohere-transcribe-arabic-07-2026 Automatic Speech Recognition (ASR) model. It has been specifically trained to improve transcription accuracy for Moroccan Darija.

Model Details

• Base Model: CohereLabs/cohere-transcribe-arabic-07-2026
• Dataset: Fine-tuned on the atlasia/DODa-audio-dataset (Moroccan Darija audio samples).
• Task: Speech-to-Text / Automatic Speech Recognition (ASR)
• Language: Moroccan Arabic (Darija)

Training Configuration

The model was fine-tuned using Parameter-Efficient Fine-Tuning (PEFT) with LoRA and 8-bit quantization.

• Steps: 500
• Batch Size: 4 (with Gradient Accumulation Steps = 4)
• Learning Rate: 1e-4
• Precision: 32-bit (fp16 disabled to prevent NaN loss in the audio architecture)

How to Use

You can use this adapter by loading the base model and merging it with this LoRA repository.

import librosa
import torch
from transformers import AutoFeatureExtractor, AutoTokenizer, CohereAsrForConditionalGeneration
from peft import PeftModel

base_model_id = "CohereLabs/cohere-transcribe-arabic-07-2026"
adapter_repo = "amzilmustapha/cohere-darija-lora-10-7-26"
hf_token = "YOUR_HF_TOKEN" # Required for the gated base model

# Load processors
feature_extractor = AutoFeatureExtractor.from_pretrained(base_model_id, trust_remote_code=True, token=hf_token)
tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True, token=hf_token)

# Load base model
base_model = CohereAsrForConditionalGeneration.from_pretrained(
    base_model_id, 
    device_map="auto", 
    trust_remote_code=True, 
    token=hf_token
)

# Apply LoRA Adapter
model = PeftModel.from_pretrained(base_model, adapter_repo, token=hf_token)

# Process Audio
audio_file = "your_darija_audio.wav" 
audio_array, sampling_rate = librosa.load(audio_file, sr=16000)
inputs = feature_extractor(audio_array, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(model.device)

# Generate Transcription
with torch.no_grad():
    generated_ids = model.generate(inputs=input_features, max_new_tokens=100)

transcription = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(transcription)