36 lines
1.7 KiB
Markdown
36 lines
1.7 KiB
Markdown
---
|
|
license: cc-by-nc-nd-4.0
|
|
language:
|
|
- en
|
|
- hi
|
|
- mr
|
|
tags:
|
|
- indic
|
|
- llama-3.2
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Akshar Nano V2 (Llama 3.2 1B) 🇮🇳
|
|
|
|
**Akshar Nano V2** is a sovereign, highly efficient 1-Billion parameter AI model developed by **Shreyas Chavan** at **Vyom AI**.
|
|
It is natively bilingual, trained to understand and reason in English, Hindi, and Marathi.
|
|
|
|
## Model Details
|
|
- **Developer:** Shreyas Chavan (Vyom AI)
|
|
- **Architecture:** 1 Billion Parameters (Llama 3.2 base)
|
|
- **Training:** Fine-tuned on a custom 25,000+ row trilingual dataset using Unsloth.
|
|
- **Languages:** English, Hindi, Marathi
|
|
- **Format:** 16-bit Safetensors (Ready for GGUF conversion)
|
|
|
|
## Capabilities
|
|
Akshar Nano V2 has been meticulously fine-tuned to excel in several key areas:
|
|
1. **Indic Mathematics & Logical Reasoning:** Capable of solving math problems and logical puzzles in Hindi and Marathi.
|
|
2. **Fluent Conversational AI:** Understands cultural nuances and can hold natural, flowing conversations in native Indic languages without sounding robotic.
|
|
3. **Identity & Sovereignty:** Deeply understands its identity as Akshar Nano, built by Vyom AI, ensuring a consistent and proud persona.
|
|
4. **General English Knowledge:** Retains the vast general knowledge of the base Llama 3.2 model while significantly boosting its Indic language capabilities.
|
|
|
|
## Usage
|
|
This model is exported in standard Hugging Face 16-bit format. It includes the proper `GenerationConfig`, `tokenizer`, and `special_tokens_map.json` to ensure it stops generating correctly and formats chat templates flawlessly.
|
|
|
|
To use this model in Ollama or llama.cpp, you can easily convert this repository to GGUF using the Hugging Face "GGUF My Repo" Space!
|