Llama-SMUGRI-7B-Instruct-MTI

An instruction-tuned version of tartuNLP/Llama-SMUGRI-7B base model continually pre-trained from meta-llama/Llama-2-7b-hf to support Võro, Komi, and Livonian. The model additionally supports English, Estonian, Finnish, and Russian, however those languages were not the focus.

The instruction-tuning dataset consists of supporting instructions in Estonian, Finnish, English, and Russian, and Alpaca-style instructions translated into Võro, Livonian, and Komi with Neurotõlge. See our paper for more details (model referenced as Llama-SMUGRI-Instruct SupInst+TrAlpaca).

Usage

We trained and evaluated our model with transformers==4.36.2.

Example usage:

from transformers import pipeline
import torch

pipe = pipeline("text-generation", model="tartuNLP/Llama-SMUGRI-7B-Instruct-MTI", torch_dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "user", "content": "Kolm nõvvo, et terveq püssüq."},
]

prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.6, top_k=50, top_p=0.9)
print(outputs[0]["generated_text"][len(prompt):])

Citation

@misc{purason2024llmsextremelylowresourcefinnougric,
      title={LLMs for Extremely Low-Resource Finno-Ugric Languages}, 
      author={Taido Purason and Hele-Andra Kuulmets and Mark Fishel},
      year={2024},
      eprint={2410.18902},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2410.18902}, 
}