PartAI/Dorna2-Llama3.1-8B-Instruct

Model Details

The Dorna models are a family of decoder-only models, specifically trained/fine-tuned on Persian data, developed by Part AI. As a new release, an 8B instruct model from this family is being made available. Dorna2-Llama3.1-8B-Instruct is built using the Meta Llama 3.1 Instruct model.

How to use

To test and use model freely on Hugging Face Spaces click here!

You can also run conversational inference using the Transformers Auto classes with the generate() function. Let's look at an example.

import torch
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM


tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system",
     "content": "You are a helpful Persian assistant. Please answer questions in the asked language."},
    {"role": "user", "content": "کاغذ A4 بزرگ تر است یا A5؟"},
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.3,
    top_p=0.85,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

You can also use the notebook below to test the model in Google Colab.

Evaluation

Comparative Evaluation

This evaluation compares Dorna2-Llama3.1-8B-Instruct, Llama3.1-8B-Instruct, and other fine-tuned Llama3.1-8B models. For broader comparisons among various large language models (LLMs), please refer to the Open Persian LLM Leaderboard, which provides a comprehensive evaluation across multiple LLMs.

Tasks and Evaluation Framework

Five specialized tasks have been carefully curated to evaluate and benchmark the models. Each task has been designed to challenge different aspects of the models' capabilities. These tasks include:

• Part Multiple Choice: Focuses on common knowledge and reasoning in a multiple-choice format.
• ARC Easy: Tests on easy-level general knowledge.
• ARC Challenge: Assesses models on harder questions requiring advanced reasoning.
• MMLU Pro: Covers professional-level exams.
• AUT Multiple Choice Persian: Specialized Persian-language examination.

Each dataset is entirely in Persian, offering a unique and robust testing ground for LLMs in non-English settings. Collectively, the datasets contain over 40k samples, spanning diverse linguistic and technical challenges such as Common Knowledge, Reasoning, Summarization, and Specialized Examinations.

Evaluation Results

Model	Average Accuracy	Part Multiple Choice	ARC Easy	ARC Challenge	MMLU Pro	AUT Multiple Choice Persian
PartAI/Dorna2-Llama3.1-8B-Instruct	50.72	34.48	79.59	64.42	21.47	53.64
O1-OPEN/OpenO1-LLama-8B-v0.1	50.22	34.66	77.87	63.08	21.24	54.24
meta-llama/Llama-3.1-8B-Instruct	50.14	36.68	78.40	60.40	21.00	54.24
NousResearch/Hermes-3-Llama-3.1-8B	48.77	35.01	77.01	58.39	21.00	52.46
Skywork/Skywork-o1-Open-Llama-3.1-8B	34.15	27.02	47.12	41.61	14.55	40.43

Contact us

If you have any questions regarding this model, you can reach us via the community on Hugging Face.