malayalam-llama3-manglish

malayalam-llama3-manglish : GGUF

This model is a fine-tuned version of Meta LLaMA 3.1 8B, optimized for Malayalam and Manglish (Malayalam written in English script) conversations.

It was fine-tuned using LoRA with Unsloth and converted to GGUF format for efficient local inference with llama.cpp.

👤 Author

Ash

🧠 Model Overview

This model is designed to:

Understand Manglish inputs Generate natural Malayalam / Manglish responses Handle casual conversational dialogue Work efficiently on low-resource systems using GGUF 📊 Dataset Details

Dataset used: https://github.com/mhdashikofficial/Manglish-LLM-dataset

Dataset Characteristics: ~20,000 conversation samples Chat-style message format (system, user, assistant) Mix of Malayalam and Manglish Human-like conversational tone Example Format: { "messages": [ {"role": "system", "content": "..."}, {"role": "user", "content": "..."}, {"role": "assistant", "content": "..."} ] } ⚙️ Training Details Base Model: Meta LLaMA 3.1 8B Fine-tuning Method: LoRA (Unsloth) Training Steps: 300 Sequence Length: 1024 Hardware: Google Colab (T4 GPU) ⚙️ Quantization Format: GGUF Method: Q4_K_M Optimized for fast inference and low memory usage 💡 Usage

Run with llama.cpp:

llama-cli -hf AlexGostroot/malayalam-llama3-manglish --jinja 📦 Available Files meta-llama-3.1-8b.Q4_K_M.gguf ⚠️ Limitations Limited reasoning depth (low training steps) May mix Malayalam and English inconsistently Not suitable for complex or critical tasks 🚀 Future Improvements Increase training steps (1000+) Add more native Malayalam data Improve response consistency Expand dataset diversity 🧾 Notes

This model is intended for:

Experimentation Local AI applications Malayalam conversational systems

Trained and converted using Unsloth: https://github.com/unslothai/unsloth