ninadp/marathi-mitra-phi3

🌸 Marathi Mitra — माझा मराठी मित्र

Fine-tuned Phi-3 Mini for Marathi vocabulary learning, built as a personalized tool to help my daughter learn Marathi.

Model Details

Property	Value
Base Model	microsoft/Phi-3-mini-4k-instruct
Fine-tuning Method	QLoRA (SFT)
LoRA Rank	r=32, alpha=64
Training Examples	30 Marathi vocabulary items
Best Experiment	exp4_lr2e4_epochs25_r32
Format Score	36.4%
Training Hardware	Google Colab T4 GPU

What It Does

Given an English word, generates a Marathi lesson with:

• Marathi word in Devanagari script
• Pronunciation guide
• Example sentence
• Fun fact for kids

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    torch_dtype=torch.float16,
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, "ninadp/marathi-mitra-phi3")
tokenizer = AutoTokenizer.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    trust_remote_code=True,
)

prompt = """### Instruction:
You are Marathi Mitra, a friendly Marathi teacher for kids.

### Input:
Teach me the Marathi word for: butterfly

### Response:
"""
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=150)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Fine-tuned using Supervised Fine-Tuning (SFT) with QLoRA on 30 Marathi vocabulary examples across 4 hyperparameter experiments.

Experiment	LR	Epochs	Loss	Score
Baseline	N/A	N/A	N/A	11.2%
Exp1	2e-4	5	1.29	12.8%
Exp2	2e-4	25	0.20	28.8%
Exp3	1e-4	25	0.37	16.0%
Exp4	2e-4	25	0.22	36.4% ✅

Limitations

• Trained on only 30 examples — vocabulary coverage is limited
• May generate incorrect Marathi words for unseen vocabulary
• Format learned well; accuracy improves with more data
• Retraining with 200+ examples planned

Live Demo

🚀 Try it on HF Spaces

GitHub

📦 Full project code