A fine-tuned version of Llama-3.2-3B-Instruct on the Agri/Koli Marathi dialect — a coastal dialect spoken by the indigenous fishing and farming communities of Maharashtra, India.

This model can understand instructions in both English and Marathi and respond authentically in the Agri/Koli dialect, using unique vocabulary and speech patterns not found in standard Marathi models.

📖 Model Details

Property	Details
Base Model	`unsloth/Llama-3.2-3B-Instruct-bnb-4bit`
Fine-tuning Method	LoRA (Low-Rank Adaptation) via Unsloth
Quantization	4-bit (bitsandbytes)
Training Platform	Kaggle (T4 GPU)
Training Steps	120
Final Loss	0.1161 (excellent convergence!)
Dataset Size	1,020+ bilingual instruction-output pairs

🗣️ What is the Agri/Koli Dialect?

The Agri/Koli dialect is spoken by two closely related indigenous communities of Maharashtra:

Agris: Farmers settled along the Konkan and Thane coastline.
Kolis: Traditional fishing communities of the Mumbai coast.

Their dialect shares Marathi roots but has distinct vocabulary, pronunciations, and grammar markers such as:

"हायेत" instead of standard "आहेत" (are/is)
"गावलेत" instead of standard "मिळाले" (got/found)
"मासले" for fish (plural form unique to the dialect)
Heavy use of coastal/nature-specific vocabulary

💬 Example Output

Instruction: How is the fishing today and how is your family? / आज मासेमारी कशी आहे आणि घरचे कसे हायेत?

Model Response:

आज मासा ला जायचं वाढ गावलेत आज, पोटांत मासले गावलेत आज. घराचं चांगले हायेत आज.

(Translation: Today we went fishing and got a good catch. The family is doing well today.)

🚀 How to Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Karan6124/AgriKoli-Llama-3.2-3B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request in the Agri/Koli Marathi dialect.

### Instruction:
{}

### Input:
{}

### Response:
{}"""

inputs = tokenizer(
    [alpaca_prompt.format("Tell me about your village.", "", "")],
    return_tensors="pt"
).to("cuda")

outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

📊 Training Details

Optimizer: AdamW 8-bit
Learning Rate: 2e-4
Batch Size: 2 (with 4 gradient accumulation steps)
LoRA Rank: 16
LoRA Alpha: 16
Sequence Length: 2048
Experiment Tracking: Weights & Biases (W&B)

📁 GitHub Repository

The full training pipeline, dataset generation scripts, and notebooks are available on GitHub:

👉 Edge-Explorer/AgriKoli-AI

The repository contains:

data/: The 1,020+ entry bilingual training dataset
scripts/: Data generation and training scripts
notebooks/: The Kaggle training notebook (agri-koli-llama-3-2-3b.ipynb)

⚠️ Limitations

This model is fine-tuned on a synthetic dataset of ~1,020 samples. Real-world dialect coverage may be limited.
Best results are achieved with short, conversational prompts.
The model is optimized for the Agri/Koli dialect specifically and may not perform well on standard Marathi tasks.

📜 License

This model is released under the MIT License.

Built with ❤️ for preserving the Agri/Koli dialect of Maharashtra. 🌾🐟