Indic LM Platform

In plain words

Kannada Llama is a Tensoic model instruction-tuned for Kannada text generation, extending foundational large models into regional dialects.

How to use it

Before you start

Prerequisites required to run the model locally

To run Kannada Llama locally on your computer, the following requirements must be met:

Install and set up Ollama.
Ensure your computer has enough RAM to run the model you plan to download.

For 3B models: Minimum 8 GB of RAM.
For 7B–8B models: Minimum 16 GB of RAM.
For 13B–32B models: Minimum 32 GB of RAM.

Quantised model: It is recommended to use 4-bit Quantized, K-measurement, Small model format. Easy to download and to run your system.
Storage: 5 to 20 GB of free SSD space to store downloaded model files.
Interface: Terminal or command line utility to run basic startup prompts.

No technical setup needed — anyone can try Kannada Llama using below mentioned steps.

Downloading the Kannada Llama GGUF file

Go to the Kannada Llama GGUF file on Hugging Face. Click the copy icon next to the repository name to copy the model path. (See image below.)

Run the model in your terminal

Open a terminal window and run the following command and paste the copied GGUF repository as mentioned below:

ollama run hf.co/PASTE_THE_COPIED_HERE

Press Enter.

Note: After pasting the repository, your command will look like this:

ollama run hf.co/Tensoic/Kan-LLaMA-7B-base

Wait for the model to load

Ollama will download and initialise the model. Once ready, you'll see the following prompt:

>>> Send a message (/? for help)

Start chatting

Type your message at the prompt and press Enter to interact with the model.

Exit the chat

To end the session, type:

/bye

Then press Enter.

Languages & scripts supported

Works in these languages, in both native script and Roman typing.

ಕನ್ನಡ KannadaEnglish English+ Roman & code-mixed

Strengths & limits

An honest look at what it does well and where it struggles.

Good at

✓Native language processing

✓Text generation

Where it struggles

!Highly complex code generation

!Low-resource dialects without fine-tuning

Technical details & licence

Training dataContinually pretrained with LoRA on ~600 million Kannada tokens from the CulturaX dataset (~$170 of A100 GPU time), then instruction fine-tuned on a Kannada-translated GPT-Teacher-style dataset.

SizeLlama-2-7B continually pretrained via LoRA on ~600M Kannada tokens (CulturaX), with vocabulary expanded from 32K to ~49,420 tokens via a merged Kannada SentencePiece tokenizer.

LicenceThe continually-pretrained base checkpoint is MIT-licensed (permissive), while the instruction-tuned SFT checkpoints are released under CC BY-NC-SA/CC BY-NC 4.0 — free for research and non-commercial use only.

Commercial costFree to self-host (open weights)

Version historyKan-LLaMA-7B-base · Dec 2023 — LoRA continually-pretrained base checkpoint, MIT licenceKan-LLaMA-7B-SFT-v0.5 · Jan 2024 — Latest instruction-tuned checkpoint, CC BY-NC 4.0

About the maker

Who builds and maintains Kannada Llama.

Tensoic

1 models

Tensoic is a small, community-style AI team known for Kan-LLaMA (Kannada Llama), a Llama-2 model continually pre-trained and fine-tuned on Kannada text and released openly on Hugging Face. It's a good example of low-budget, open fine-tuning making a real difference for an under-served language.

View their models →See all organisations →