---
license: mit
datasets:
- Devavrat28/All-about-Peshwes
language:
- mr
base_model:
- unsloth/gemma-3-12b-it-unsloth-bnb-4bit
tags:
- history
- AI
- culture
- Fine-tuning
- Gemma3
---
```markdown
📜 Peshwai Historian AI (Marathi Language Model)
An LLM fine-tuned to answer deep, culturally rich questions about the Peshwa era of Pune — in fluent Marathi.

✨ Overview

`peshwai-historian-ai` is a fine-tuned language model focused on the **Maratha Peshwa dynasty, trained on **curated historical text in Marathi. The model can answer questions about lesser-known events, policies, and figures from the 18th-century Pune region with contextual, factual, and culturally accurate responses — especially useful for educators, students, historians, and heritage lovers.


🔍 What Makes It Special?

📖 Marathi-native output: Generates grammatically rich and natural Marathi text
🕰️ Historical awareness: Avoids commonly known facts (like Bajirao, Shaniwarwada) and focuses on underrepresented topics
🧠 Fine-tuned on real historical documents: Including rare facts about Nana Phadnavis, Mahadji Shinde, diplomacy, cultural shifts, etc.
🗣️ Few-shot prompting optimized: Learns from examples to improve answer quality


🧠 Example Prompt

```marathi
विषय: नाना फडणवीसांचे गुप्त राजकारण
सविस्तर माहिती: 
```

### 💬 Sample Response

```marathi
नाना फडणवीस हे पेशवाईतील अत्यंत मुत्सद्दी आणि धोरणशक्ती असलेले व्यक्तिमत्व होते. माधवराव पेशव्यांच्या मृत्यूनंतर, सत्तेची रिकामी जागा भरून काढण्यासाठी त्यांनी 'बारभाई मंडळ' तयार केले...
```

---

## 🧪 How to Use

### In Python:
```python
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Devavrat28/peshwai-historian-ai",  # replace with your HF username
    max_seq_length = 4096,
    dtype = torch.float16,
    load_in_4bit = True
)

FastLanguageModel.for_inference(model)

prompt = "विषय: माधवराव पेशव्यांचा आरोग्यावर झालेला परिणाम\nसविस्तर माहिती:"
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

---

## 💡 Fine-Tuning Details

| Setting              | Value                          |
|----------------------|--------------------------------|
| Base Model           | Gemma / LLaMA / Open LLM       |
| Fine-tuning Method   | Supervised Fine-Tuning (SFT)   |
| Framework            | [Unsloth](https://github.com/unslothai/unsloth) + 🤗 TRL |
| Dataset              | Marathi historical text (~50K tokens) |
| Technique            | Continued pretraining + SFT    |
| Language             | मराठी (Marathi)                |

---

## 🛠️ Intended Use

- 📚 **Educational apps** in schools or colleges
- 🏛️ Museums or digital history archives
- 🗣️ Voice-based Marathi chatbots for local history
- 📖 Research tools for historians and scholars

---

## 📜 Citation

If you use this model in research or production, please consider citing:

```
@misc{peshwaiHistorian2024,
  title={Peshwai Historian AI: A Marathi LLM for Regional Heritage},
  author={Devavrat Samak},
  year={2024},
  howpublished={\url{https://huggingface.co/devavrat/peshwai-historian-ai}},
}
```

---

## ❤️ Credits

Developed by [Devavrat Samak](https://huggingface.co/devavrat)  
Inspired by the rich cultural heritage of Pune and the legacy of the Peshwas.

---

## 📬 Feedback / Contributions

I welcome pull requests, prompts, dataset contributions, and collaborations. Reach out via Hugging Face or GitHub.