ParsBench provides toolkits for benchmarking Large Language Models (LLMs) based on the Persian language. It includes various tasks for evaluating LLMs on different topics, benchmarking tools to compare multiple models and rank them, and an easy, fully customizable API for developers to create custom models, tasks, scores, and benchmarks.

• Variety of Tasks: Evaluate LLMs across various topics.
• Benchmarking Tools: Compare and rank multiple models.
• Customizable API: Create custom models, tasks, scores, and benchmarks with ease.

I was trying to fine-tune an open-source LLM for the Persian language. I needed some evaluation to test the performance and utility of my LLM. It leads me to research and find this paper. It's great work that they prepared some datasets and evaluation methods to test on ChatGPT. They even shared their code in this repository.

So, I thought that I should build a handy framework that includes various tasks and datasets for evaluating LLMs based on the Persian language. I used some parts of their work (Datasets, Metrics, Basic prompt templates) in this library.

Install Math Equivalence package manually:

pip install git https://github.com/hendrycks/math.git

Install ParsBench using pip:

pip install parsbench

Load the pre-trained model and tokenizer from the HuggingFace and then, evaluate the model using the PersianMath task:

from transformers import AutoModelForCausalLM, AutoTokenizer

from parsbench.models import PreTrainedTransformerModel
from parsbench.tasks import PersianMath

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2-72B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-72B-Instruct")

tf_model = PreTrainedTransformerModel(model=model, tokenizer=tokenizer)

with PersianMath() as task:
    results = task.evaluate(tf_model)

For example, we run our local models using Ollama:

ollama run qwen2
ollama run aya

Then we benchmark those models using the ParsBench.

from parsbench.benchmarks import CustomBenchmark
from parsbench.models import OpenAIModel
from parsbench.tasks import ParsiNLUMultipleChoice, PersianMath, ParsiNLUReadingComprehension

qwen2_model = OpenAIModel(
    api_base_url="http://localhost:11434/v1/",
    api_secret_key="ollama",
    model="qwen2:latest",
)
aya_model = OpenAIModel(
    api_base_url="http://localhost:11434/v1/",
    api_secret_key="ollama",
    model="aya:latest",
)

benchmark = CustomBenchmark(
    models=[qwen2_model, aya_model],
    tasks=[
        ParsiNLUMultipleChoice,
        ParsiNLUReadingComprehension,
        PersianMath,
    ],
)
result = benchmark.run(
    prompt_lang="fa",
    prompt_shots=[0, 3],
    n_first=100,
    sort_by_score=True,
)
result.show_radar_plot()

You can import the class of above tasks from parsbench.tasks and use it for evaluating your model.

• Benchmark Aya models:
• Benchmark Ava models:
• Benchmark Dorna models:
• Benchmark MaralGPT models:

Here are the names of companies/people who helped us to keep maintaining this project. If you want to donate this project, see this page.

• AvalAI: They gave us free OpenAI API credit several times in their "AvalAward" program. It helped us for doing R&D and benchmarking GPT models.
• Basalam: They voluntarily helped us to run the benchmarks on open-weight models and build the ParsBench Leaderboard.

Contributions are welcome! Please refer to the contribution guidelines for more information on how to contribute.

ParsBench is distributed under the Apache-2.0 license.

For support or questions, please contact: [email protected] Feel free to let me know if there are any additional details or changes you'd like to make!