Luau Bench

An evaluation harness for benchmarking LLM-generated Luau code. Highly-inspired from EleutherAI/lm-evaluation-harness.

Installation

git clone https://github.com/imezx/luau-bench.git
cd luau-bench

# With all optional providers & datasets
pip install -e ".[all]"

# For specific providers
pip install -e ".[anthropic]"  # Claude models
pip install -e ".[datasets]"   # HuggingFace dataset loader
pip install -e ".[progress]"   # tqdm progress bars for long runs

# through pip
pip install luau-bench --no-cache-dir

Runtime Setup

Luau Bench requires a Luau runtime for executing generated code. Install one or more with the bundled script:

# Install everything (Luau CLI + Zune + StyLua)
./tools/install-runtime.sh --all

# Install Zune only (recommended runtime)
./tools/install-runtime.sh --zune

# Install the Luau CLI (also installs luau-analyze for static diagnostics)
./tools/install-runtime.sh --luau

# verify the environment
luau-bench info

Supported runtimes in priority order: luau, zune. The LUAU_RUNTIME environment variable can pin a specific binary path.

Quick Start

mkdir my_tasks
cp templates/task_basic.yaml my_tasks/

luau-bench validate --include-path ./my_tasks

luau-bench selftest --include-path ./my_tasks

# Run a benchmark - results printed to terminal only
luau-bench run --provider ollama --model codellama \
    --include-path ./my_tasks

# Run & export results (all formats)
luau-bench run --provider ollama --model codellama \
    --include-path ./my_tasks --export

# Run & export specific formats
luau-bench run --provider ollama --model codellama \
    --include-path ./my_tasks --export html,json

CLI Reference

luau-bench run          Evaluate a model against tasks
luau-bench ls           List available tasks & groups
luau-bench validate     Check task YAML files for errors
luau-bench selftest     Run reference solutions against test harnesses
luau-bench compare      Compare two or more result JSON files
luau-bench cache        Manage the generation result cache
luau-bench info         Show environment info

run

luau-bench run \
    --provider ollama \
    --model codellama \
    --include-path ./my_tasks \
    --tasks task1,task2 \
    --num-samples 5 \
    --parallel 4 \
    --task-parallel 4 \
    --export html,json,md \
    --output ./results \
    --log-samples --show-samples \
    --temperature 0.0 \
    --max-tokens 4096 \
    -v

Results are printed to the terminal by default. Pass --export to write files. Use --export alone to export all formats (json, md, html), or --export html,json to select specific ones. The --output directory is only used when --export is active.

Supported Providers

Provider	--provider value	Notes
vLLM	vllm	Full support including loglikelihood
OpenAI	openai	Set OPENAI_API_KEY
Anthropic Claude	anthropic	Set ANTHROPIC_API_KEY; no loglikelihood support
Ollama	ollama	Local models via ollama serve
TGI	tgi	HuggingFace Text Generation Inference
LM Studio	lmstudio	Local via LM Studio server

Loglikelihood scoring (output_type: loglikelihood) requires a backend that exposes the /v1/completions endpoint with echo=True & logprobs=1. This includes vLLM, TGI, & llama.cpp. It is not supported by the Anthropic adapter.

Contributing

We accept any well-formed task that has a clear purpose, correct reference solution, & sufficient test coverage.