Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

google-ai-edge/eval

9PythonApache-2.0

Google releases a CLI tool that benchmarks on-device AI models the same way as regular ones

ai-edge-eval is a command-line evaluation tool from Google that lets you test both LiteRT LM models, Google's lightweight on-device model format, and standard models like HuggingFace checkpoints using the same interface. It runs text and vision+text benchmarks such as ifeval, mmlu, piqa and mmmu through existing evaluation frameworks lm-eval and lighteval. Users can also register their own datasets and scoring functions to build custom evaluation tasks.

What it does

  1. Provides one CLI that evaluates both LiteRT LM models and native models like HuggingFace under a shared workflow
  2. Supports two backend frameworks, lm-eval and lighteval, covering both text-only and multimodal (vision + text) tasks
  3. Offers --limit and --sample-range flags to run quick smoke tests on a small slice of data instead of the full benchmark
  4. Lets users define custom tasks by supplying a JSON Lines dataset and a Python scoring function via a registration file
  5. Runs on Linux, macOS, and WSL2, and is released under the Apache 2.0 license

Why it matters

It gives developers building small on-device language models a standardized way to measure performance across different model formats. It also flags that each benchmark dataset carries its own license, so users must check usage terms themselves.

Terms in this repo

  • LiteRT LM · Google's lightweight model format built for running on-device
  • lm-eval · A benchmark evaluation harness for language models made by EleutherAI
  • lighteval · An alternative evaluation framework from the HuggingFace ecosystem
  • multimodal · Handling both images and text together as input
  • JSON Lines (.jsonl) · A file format storing one JSON object per line

Repository description (English)

google-ai-edge/eval

Owner · google-deepmind

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA