# carlini/yet-another-applied-llm-benchmark

A benchmark to evaluate language models on questions I've previously asked them to solve.

Repository: https://github.com/carlini/yet-another-applied-llm-benchmark
Canonical: https://ross.abutalabs.com/products/yet-another-applied-llm-benchmark
Language: Python
License: GPL-3.0
License Family: copyleft
Last push: 2025-04-27T06:25:09+00:00

## Health v2 (maintenance only)
Score: 34/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 18, release rhythm 35, longevity 70
- inputs: {"age_days": 990, "days_push": 493, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1064, forks 77 (observed 2026-08-28T04:03:26.418265+00:00)

## What it is
A personal benchmark of nearly 100 applied tests for evaluating how well large language models perform on practical tasks the author has actually asked LLMs to solve, built on a small dataflow DSL for defining test pipelines. Tests run generated code in Docker containers and use programmatic and LLM-based evaluators to score model outputs.

## Use cases
- benchmark large language models on real-world applied tasks
- compare how different LLMs perform on code generation and reasoning tests
- evaluate whether a model can write and run working programs in python, c, and bash
- add custom LLM evaluation tests using a simple dataflow DSL
- check if a model can parse grammars, convert english to sql, or decompile bytecode
- track LLM capability regressions across model versions

## When to choose
- you want applied, task-based evaluation of LLMs rather than academic question answering
- you need a lightweight harness to run generated code in containers and automatically judge outputs
- you want to add your own realistic test cases quickly with a minimal DSL
- you want reference scores for popular models like GPT-4o, Claude, and Gemini on practical tasks

## When to avoid
- you need a rigorously validated academic benchmark with statistical guarantees
- you require broad standardized coverage aligned with published leaderboards
- you want a benchmark focused on a single narrow capability rather than diverse applied tasks
- you cannot run Docker or execute generated code locally

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, llm-inference, agent-framework, testing, developer-tools
- domain: large-language-models, machine-learning, developer-tools, testing, artificial-intelligence
- platform: python, cli
- tags: llm-benchmark, evaluation-harness, dsl, prompt-evaluation, code-generation-evaluation, model-evaluation, applied-tasks, docker

## Member repositories
- carlini/yet-another-applied-llm-benchmark (main) score 34

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:26.418265+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:56:03.300533+00:00, confidence not recorded.
  - readme: https://github.com/carlini/yet-another-applied-llm-benchmark (fetched 2026-08-28T04:03:26.418265+00:00, sha 05b6ebc1f6dd)
- Data as of 2026-08-30T08:39:29.467469+00:00.