README.md

---
base_model: Lo-Renz-O/Mistral-7B-CPT-Malagasy-v2-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- mistral
license: apache-2.0
language:
- en
datasets:
- Lo-Renz-O/alpaca-gpt4-MG
---
## Model Description

This model is an **Instruction Fine-Tuned** adaptation of  
[**Mistral-7B-CPT-Malagasy-v2-bnb-4bit**](https://huggingface.co/Lo-Renz-O/Mistral-7B-CPT-Malagasy-v2-bnb-4bit), optimized to **follow instructions in the Malagasy language**.

It is designed to serve as a **Malagasy AI assistant** with ability to:
- Follow instructions  
- Answer questions and explain concepts in Malagasy  
- Structure outputs (steps, lists, comparisons)
- Perform everyday reasoning tasks

This version is specifically meant for **chat & assistant usage**.

---

## Intended Uses & Limitations

✅ **Recommended Use Cases**
- RAG (Retrivial Augmented Generation)
- Malagasy conversational AI assistant
- Task-oriented instructions (How-to, Q&A, explanations)
- Educational and government chatbots
- Research on Malagasy instruction-following LLMs
- Cultural and contextual language research
  
⚠️ **Limitations**
- May hallucinate if asked for unknown facts
- Not aligned

---

## Training Details

- **Base Model:** Lo-Renz-O/Mistral-7B-CPT-Malagasy-v2-bnb-4bit  
- **Method:** Instruction Fine-Tuning (SFT) with LoRA adapters  
- **Dataset:** [Malagasy instruction dataset](https://huggingface.co/datasets/Lo-Renz-O/alpaca-gpt4-MG)
- **Hardware:** 1 × GPU T4 
- **Number of Epochs:** 1  
- **Training Time:** ~65 hours  
- **Objective:** Improve task-following in Malagasy

**Training Loss Curve:**  
<p align="left">
  <img src=https://cdn-uploads.huggingface.co/production/uploads/67a3188ac7d276b8df08d11f/5vjekdGyRs3NKEHe9k6Qr.png alt="Training Loss Curve" width="900"/>
</p>

---

## Inference Example Usage
**code**
```python
from unsloth import FastLanguageModel
from transformers import TextStreamer

model_name = "Lo-Renz-O/Mistral-7B-instruct-Malagasy-bnb-4bit"
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

conversation_history = []

system_message = (
    "Ianao dia mpanampy manampahaizana sy mahafantatra tsara izay mamaly "
    "amin'ny teny Malagasy. Valio amin'ny fomba mazava, am-pahatsorana, "
    "ary omeo fanazavana feno raha ilaina. Ataovy voalamina tsara ny valiny, "
    "mampiasa lohateny, lohateny kely, sy teboka miavaka rehefa ilaina."
)

def build_messages(history, user_message):
    messages = [{"role": "system", "content": system_message}]
    for turn in history:
        messages.append({"role": "user", "content": turn["user"]})
        messages.append({"role": "assistant", "content": turn["assistant"]})
    messages.append({"role": "user", "content": user_message})
    return messages

print("*-----------------Ampidiro ny fanontaniana-----------------*")

while True:
    user_input = input("User: ")
    if user_input.lower() in ["exit", "quit"]:
        break

    messages = build_messages(conversation_history, user_input)
    prompt = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
    )

    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    streamer = TextStreamer(tokenizer, skip_special_tokens=True, skip_prompt=True)

    print("\nAssistant:", end=" ", flush=True)
    output = model.generate(
        **inputs,
        max_new_tokens=1024,
        temperature=0.7,
        top_p=0.9,
        top_k=50,
        repetition_penalty=1.15,
        do_sample=True,
        eos_token_id=tokenizer.eos_token_id,
        streamer=streamer,
    )

    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
    assistant_response = generated_text.split("Assistant:")[-1].strip()

    conversation_history.append(
        {"user": user_input, "assistant": assistant_response}
    )
    print()
```
**output**
```
*-----------------Ampidiro ny fanontaniana-----------------*
User:  Inona no paikady hanampiana ny fampandrosoana any ambanivohitra?

Assistant: Ny fampandrosoana any ambanivohitra dia manondro ny fivoaran'ny faritra ivelan'ny tanàn-dehibe, ary misy paikady maromaro azo ampiharina mba hanatsarana izany.
Ny sasany amin'ireo paikady fototra indrindra dia ahitana:  1. Fampiroboroboana ny fampandrosoana maharitra: Fampiroboroboana ny fampandrosoana maharitra ny faritra
ambanivohitra izay manantitrantitra ny fiarovana ny tontolo iainana, ny fiarovana ny kolontsaina ary ny famoronana asa maharitra. Izany dia azo atao amin'ny alalan'ny
fanomezana tolotra sy fotodrafitrasa ilaina toy ny fitateram-bahoaka, toeram-pambolena sy fambolena maharitra, sehatry ny angovo azo havaozina, ary sehatry ny teknolojia
maharitra.  2. Fanohanana ny orinasa eo an-toerana: Manampy amin'ny fampihenana ny fahasahiranan'ny orinasa madinika eo an-toerana ny fanohanana ny fampivoarana azy ireo.
Ny fanomezana tontolo iainana tsara ho an'ny fandraharahana, toy ny fidirana amin'ny fampindramam-bola, ny fanofanana ary ny fanampiana, dia afaka manampy amin'ny
fampiroboroboana ny faharetan'ny orinasa madinika, ka mahatonga azy ireo ho maharitra kokoa.  3. Fanomezana fahafahana ho an'ny asa: Manampy amin'ny famoronana asa sy
fampihenana ny fahantrana ny fanomezana asa, indrindra any
```
# Uploaded finetuned  model
- **Developed by:** Lo-Renz-O
- **License:** apache-2.0
- **Finetuned from model :** Lo-Renz-O/Mistral-7B-CPT-Malagasy-v2-bnb-4bit

This mistral model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)