arbib/darija-qwen2.5-unsloth-lora_v2

Darija Qwen2.5-1.5B — LoRA Adapter (Unsloth)

Modele fine-tune sur Qwen/Qwen2.5-1.5B-Instruct pour le dialecte marocain (Darija). Entrainement via Unsloth (2x plus rapide, 40% moins de VRAM).

Dataset

• MBZUAI-Paris/Darija-SFT-Mixture
• 15 000 samples selectionnes et nettoyes

Configuration LoRA

• lora_rank: 8
• lora_alpha: 16
• lora_dropout: 0.05
• lora_target: q, k, v, o, gate, up, down

Utilisation

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name   = "/kaggle/working/darija_qwen_lora_adapter",
    max_seq_length = 512,
    dtype        = torch.float16,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)