qLMM-Lua-Eval-Pipeline

Script and data for the manuscript titled "Evaluating Quantized Large Language Models for Code Generation on Low-Resource Language Benchmarks".

Files

• ".env" contains variables used by the python scripts. Make sure to set the "MODEL_REP_PATH" variable with a path to a directory that contains the locally stored code LLMs.
• "config.json" provides the list of used code LLMs and benchmarks.
• "hfDatasetDownloader.py" downloads and formats the MultiPL-HumanEval, MultiPL-MBPP, and MCEVAL benchmarks from HuggingFace. The download benchmarks are stored inside the "benchmarks" directory.
• "genPipe.py" script that loads code LLMs one by one and applies code generation tasks to them. All generated code is stored inside the "genOutput" directory.
• "evalPipe.py" evaluates the Lua and Python code generated by the code LLMs using several metrics mentioned in the manuscript. The evaluation results are stored inside the "evalOutput" directory.
• "analysis.R" to analyze the content of the "evalOutput" directory.
• "genOutput.zip" contains the "genOutput" directory with the data from code generation.
• "evalOutput.zip" contains the "evalOutput" directory with the data from evaluation.

config.json structure

{
"name": "Unique name for the model.",
"family": "Model family name. Same at all quantization precisions.",
"id": "HuggingFace URI of the model.",
"max_tokens": "Maximum number of tokens to generate.",
"temp": "Temperature at which the model is run.",
"top_k": "Next token sampling rate.",
"lua_eos": "End-Of-Sequence tokens for Lua tasks",
"py_eos": "End-Of-Sequence tokens for Python tasks",
"qBits": "Precision. 2, 4, 8 for integer quantization precision and 16 for half-precision floating point.",
"passK": "Value of k in pass@k, For example, 10 for pass@10.",
"skip": "0 or 1. If 1 the model will be ignored by genPipe.py."
}

{
"name": "benchmark name",
"id": "jsonl file with the benchmark",
"sample": "None or integer number. If integer number N then a random sample with N tasks will be used to evaluate the models",
"skip": "0 or 1. If 1 the benchmark will be ignored by genPipe.py."
}

Setting the environment

The Python script requires following packages:

• dotenv
• pathlib
• airium
• pandas
• llama-cpp-python

Download the code LLMs and store them locally. The list of code LLMs is available inside the config.json file. Inside the .env file, make sure to set the "MODEL_REP_PATH" variable to the directory path that contains the locally stored code LLMs.

Before running this script ensure that

• Lua 3.5 or higher is installed
• Path to Lua is added to the environmental variable
• the LuaUnit package is installed
• install the CLOC tool to count lines of code (winget install AlDanial.Cloc): https://github.com/AlDanial/cloc?tab=readme-ov-file

Order of execution

1. hfDatasetDownloader.py
2. genPipe.py
3. evalPipe.py
4. analysis.R