35 lines
1.2 KiB
Markdown
35 lines
1.2 KiB
Markdown
|
---
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
license: cc-by-nc-nd-4.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- hi
|
||
|
|
- mr
|
||
|
|
tags:
|
||
|
|
- indic
|
||
|
|
- unsloth
|
||
|
|
- llama-3.2
|
||
|
|
- akshar-nano
|
||
|
|
---
|
||
|
|
# 🌟 Akshar Nano
|
||
|
|
|
||
|
|
**Designed by:** Vyom AI
|
||
|
|
**Developed by:** Shreyas Chavan
|
||
|
|
|
||
|
|
**Akshar Nano** is a sovereign, highly efficient 1-Billion parameter AI model built on Meta's Llama 3.2 architecture. It has been meticulously fine-tuned using dataset streaming on Kaggle to understand and generate culturally and linguistically precise text in **English, Hindi, and Marathi**.
|
||
|
|
|
||
|
|
## 🚀 Model Details
|
||
|
|
- **Base Model:** `unsloth/Llama-3.2-1B-Instruct`
|
||
|
|
- **Parameters:** 1 Billion
|
||
|
|
- **Quantization:** 4-bit (BitsAndBytes)
|
||
|
|
- **Training:** 4-hour Streaming SFT on NVIDIA T4
|
||
|
|
|
||
|
|
## 📚 Datasets Used
|
||
|
|
- `sarvamai/gsm8k-indic` (Marathi, English, Hindi)
|
||
|
|
- `sarvamai/samvaad-hi-v1` (Hindi & English conversations)
|
||
|
|
|
||
|
|
## ⚖️ Legal Notice & License
|
||
|
|
This model is licensed under **Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)**.
|
||
|
|
|
||
|
|
**STRICT WARNING:** It is strictly prohibited to copy, rebrand, or change the identification/persona of this model. Any unauthorized alteration of the model's identity, creator attribution, or origin will result in immediate legal action.
|