dunebench 0.3

dunebench – a lightweight evaluation tool for llama.cpp models

DuneBench

dunebench is a lightweight, local benchmarking tool for GGUF models. It allows you to evaluate Large Language Models (LLMs) across a variety of domains—including logic, coding, math, and common sense—using llama-cpp-python.

Installation

Install dunebench with pip

    pip install dunebench

or install EXE with this link

Features

• Local Evaluation: Runs entirely on your machine using GGUF models.
• GPU Accelerated: Offload layers to your GPU for faster testing.
• Multi-Domain Support: Includes 8 distinct benchmarks (Math, Coding, Science, etc.).

Usage/Examples

dunebench --model "path/to/model.gguf" --task science --limit 20

Arguments

Argument	Description	Default
--model	Path to your .gguf model file	Required
--task	The benchmark task to run	Required
--limit	Number of samples to test	10

tasks

Task Name	Dataset Used	Domain	Type
science	ai2_arc (Challenge)	Scientific Reasoning	Multiple Choice
math	gsm8k	Math Word Problems	Generation
programming	mbpp (Sanitized)	Python Coding	Code Generation
physical_logic	piqa	Physical Commonsense	Multiple Choice
common_sense	openbookqa	General Knowledge	Multiple Choice
logic	winogrande	Ambiguity Resolution	Multiple Choice
grammar	glue (CoLA)	Linguistic Acceptability	Multiple Choice
nlp	hellaswag	Sentence Completion	Multiple Choice

License

MIT

Authors

• @Rudransh Joshi