---
license: cc0-1.0
configs:
- config_name: results
  data_files:
  - split: test
    path: model_scores_2026-08-27.json
language:
- en
tags:
- benchmark-results
- language-model-evaluation
- small-language-models
---

![Veyra Banner](https://cdn-uploads.huggingface.co/production/uploads/6857f2cfae68b377f17aff8c/MfZTOclKIpOM1RbUlnfcr.jpeg)

# Micro-Model-Bench

Micro-Model-Bench is a collection of benchmark results that have been collected using [lm-evaluation-harness](https://github.com/eleutherai/lm-evaluation-harness)

### Models Recorded

2026-08-27:
- 154 model records across 49 organizations
- 125 records marked isValid: true, all other models haven't been evaluated because of gated access or lm-eval not being able to benchmark them.

### What benchmarks are included:

- ARC-Easy: arc_easy_acc, arc_easy_acc_norm
- ARC-Challenge: arc_challenge_acc, arc_challenge_acc_norm
- HellaSwag: hellaswag_acc, hellaswag_acc_norm
- PIQA: piqa_acc, piqa_acc_norm
- SciQ: sciq_acc, sciq_acc_norm
- WinoGrande: winogrande_acc
- OpenBookQA: openbookqa_acc, openbookqa_acc_norm
- BoolQ: boolq_acc
- LAMBADA: lambada_acc
- LogiQA: logiqa_acc
- MathQA: mathqa_acc

## Uses

This dataset has been made for people to be able to compare models under 150M against eachother and provide it in a fully open way without any restrictions. You can make leaderboards, visualisations, etc.

## Dataset Structure

The JSON file contains two top-level objects:

<pre><code>{
  "orgs": {
    "example-org": {
      "name": "example-org",
      "url": "https://huggingface.co/example-org"
    }
  },
  "models": [
    {
      "name": "Example-10M-Base",
      "org": "example-org",
      "params": 10000000,
      "links": {
        "card": "https://huggingface.co/example-org/Example-10M-Base"
      },
      "isCustom": false,
      "isLooped": false,
      "loopCount": 1,
      "firstCommit": "2026-08-27",
      "arc_easy_acc": 0.0,
      "arc_easy_acc_norm": 0.0,
      "isValid": true
    }
  ]
}
</code></pre>

### License

The results dataset is dedicated to the public domain under CC0 1.0. Individual models retain their own licenses and terms.

### Attribution

No attribution is required but if you are using Micro-Model-Bench for any public project we would appreciate a citation or link:

> https://huggingface.co/datasets/veyra-ai/Micro-Model-Bench

Attribution does not imply endorsement by Veyra AI or the model authors represented in this dataset.