README.md

---
license: mit
tags:
- gemma
- gguf
- llama.cpp
- indic
---

# GGUF for Indic-gemma-2b-finetuned-sft-Navarasa

This model from [Telugu-LLM-Labs](https://huggingface.co/Telugu-LLM-Labs/Indic-gemma-2b-finetuned-sft-Navarasa) is based on google/gemma-2b and has been LoRA finetuned on 9 Indian languages and English instruction datasets


```bash
git clone https://huggingface.co/iAkashPaul/Indic-gemma-2b-finetuned-sft-Navarasa-GGUF # & cd into it, update paths accordingly

# build llama.cpp for your hardware https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#build

./main --file prompt.md --lora ./models/ggml-adapter-model.bin --lora-base ./models/indic-llm_Q8.gguf

./main --file prompt.md -m ./models/merged_indic_llm_Q8.gguf -ngl 99

```

## Prompt template for Instruction adherence-

Save this to a file(ex. prompt.md) & load it with the main executable.

```markdown
### Instruction: Translate following sentence to Kannada.

### Input: This model is developed by Telugu LLM Labs

## Response: 
```

## Performance

* LORA+BASE (not merged)
  
  * ```
    ./server --lora ./models/ggml-adapter-model.bin --lora-base ./models/indic-llm_Q8.gguf -m ./models/indic-llm_Q8.gguf  
    ```
    
  * ![](indic-llm-q8.jpg)

* Merged model

  * ```
    ./server -ngl 20  -m ./models/merged_indic_llm_Q8.gguf 
    ```
 
  * ![](Q8-75tok.png)