viplismism/Qwen2.5-7B-Mewari-MLX-LoRA

Qwen2.5-7B Mewari Translation (MLX LoRA)

A LoRA fine-tuned adapter for English to Mewari translation, built on top of Qwen/Qwen2.5-7B-Instruct using MLX on Apple Silicon.

Mewari (मेवाड़ी) is a Rajasthani language spoken in the Mewar region of Rajasthan, India, written in Devanagari script.

Usage

With MLX (Apple Silicon)

from huggingface_hub import snapshot_download
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

# Download adapter from HuggingFace
adapter_path = snapshot_download(repo_id="viplismism/Qwen2.5-7B-Mewari-MLX-LoRA")

# Load base model with adapter
model, tokenizer = load("Qwen/Qwen2.5-7B-Instruct", adapter_path=adapter_path)

messages = [
    {"role": "system", "content": "You are an expert translator specializing in English to Mewari translation. Provide only the direct Mewari translation in Devanagari script, nothing else."},
    {"role": "user", "content": 'English text to translate: "Hello, how are you?"\n\nProvide the Mewari translation:'},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
sampler = make_sampler(temp=0.7, top_p=0.9)
response = generate(model, tokenizer, prompt=prompt, max_tokens=256, sampler=sampler)
print(response)
# Output: नमस्ते, थूं कैसै है?

Example Translations

English	Mewari
Hello, how are you?	नमस्ते, थूं कैसै है?
The weather is very hot today	आज मौसम घणो गरम है।
Please sit down and have some tea	कृपया बैठ जावो अर कुछ चाय खावो।
What is your name?	थारो नाव क्या है?
Where are you going tomorrow?	काल थें कठै जावां?
My children go to school every day	म्हारै बाचेरे हर दिन स्कूल जावै है।

Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-7B-Instruct
Method	LoRA (MLX)
Training Data	2,700 English-Mewari pairs
Validation Data	300 English-Mewari pairs
LoRA Rank	64
LoRA Alpha	128
LoRA Dropout	0.1
Learning Rate	1e-5
Batch Size	1
Iterations	1000
Max Seq Length	512
Grad Checkpoint	Yes

Training Results

Metric	Value
Train Loss	2.076 → 0.265
Val Loss	2.703 → 0.290
Best Val Loss	0.282 (iter 700)
Peak Memory	18.856 GB
Hardware	Apple M4 Max (36GB)

Limitations

• Optimized for simple to moderate sentence translation
• May produce repetition on certain complex or compound sentences
• Best used with temperature 0.7 and top_p 0.9
• MLX adapter format — designed for Apple Silicon inference