SKILL.md
AI Model Benchmarking Guide
Overview
Rigorous evaluation is the backbone of machine learning research. A model is only as credible as its evaluation protocol: which benchmarks were used, how metrics were computed, whether results are reproducible, and how they compare to baselines. The proliferation of LLMs has made this both more important and more complex, with over 60 established benchmarks and a rapidly evolving landscape.
This guide covers the practical side of model benchmarking: how to use the EleutherAI Language Model Evaluation Harness (lm-evaluation-harness), how to select benchmarks for different research claims, how to avoid common evaluation pitfalls, and how to present results for publication. The focus is on academic rigor rather than leaderboard chasing.
Whether you are evaluating a fine-tuned model for a paper, comparing architectures for an ablation study, or reviewing a submitted manuscript's evaluation section, these patterns will help ensure the evaluation is sound.
The lm-evaluation-harness
The EleutherAI lm-evaluation-harness is the de facto standard for LLM evaluation in academic research, supporting 60+ tasks and used by most major LLM papers.
Installation and Basic Usage
# Install
pip install lm-eval
# Run a single benchmark
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu \
--batch_size auto \
--output_path results/llama2-7b/
# Run multiple benchmarks
lm_eval --model hf \
--model_args pretrained=meta-llama/Llama-2-7b-hf \
--tasks mmlu,hellaswag,arc_challenge,winogrande,truthfulqa_mc2 \
--batch_size auto \
--num_fewshot 5 \
--output_path results/llama2-7b/
Programmatic API
import lm_eval
results = lm_eval.simple_evaluate(
model="hf",
model_args="pretrained=meta-llama/Llama-2-7b-hf",
tasks=["mmlu", "hellaswag", "arc_challenge"],
num_fewshot=5,
batch_size="auto",
device="cuda",
)
# Access results
for task, metrics in results["results"].items():
print(f"{task}: {metrics}")