clmi / README.md
ucr-max's picture
Deploy prebuilt CompactLMIndex
7233de0 verified
|
Raw
History Blame Contribute Delete
2.53 kB
metadata
title: Compact Language Model Index
emoji: 📊
colorFrom: indigo
colorTo: blue
sdk: static
app_file: build/index.html
fullWidth: true
short_description: Chance-normalized evaluation of compact language models.

Compact Language Model Index

CompactLMIndex is a curated benchmark for smaller language models. It uses one fixed protocol to compare models across commonsense reasoning, science and world knowledge, and grammatical competence.

Evaluation suite

The published composite includes six benchmarks:

Area Benchmark
Situational commonsense HellaSwag
Physical commonsense PIQA
Science and world knowledge ARC-Easy
Science and world knowledge ARC-Challenge
Conceptual commonsense CommonsenseQA
Grammatical competence BLiMP

All models use zero-shot, non-chat log-likelihood prompts. A complete run of the core suite is required for a model to receive a rank.

Scoring

Each benchmark score is normalized for random chance:

100 × (accuracy − chance) / (1 − chance)

Random performance maps to 0 and perfect performance maps to 100. Scores below chance remain negative.

The composite is domain-balanced:

  • commonsense: 50%, split equally between HellaSwag, PIQA, and CommonsenseQA;
  • science and world knowledge: 25%, with ARC-Easy contributing 75% and ARC-Challenge 25% within that domain;
  • grammatical competence: 25%, from BLiMP.

BLiMP is macro-averaged across its 12 linguistic categories. CommonsenseQA scores the full answer text with length normalization rather than answer labels such as A–E.

Uncertainty and ranking

Every displayed benchmark and composite interval is a 95% percentile bootstrap interval. Model comparisons use paired bootstrap resampling on matched evaluation items. Rank ranges show when neighbouring models cannot be reliably distinguished under those comparisons.

Scope and inclusion

CompactLMIndex is curated rather than an open intake list. Published results are checked against their saved evaluation artifacts, and final inclusion and presentation decisions remain with Universal Computing Research.

If you have results for a model that fits the benchmark scope, open a discussion on this Space for consideration.

Credits

Inspired by the Open LLM Leaderboard, Axiomic Labs Open SLM Leaderboard, and OpenAI Parameter Golf.