---
license: unknown
task_categories:
  - question-answering
  - text-generation
tags:
  - llm-as-judge
  - benchmark
  - gpqa
  - mmlu-pro
  - jury-system
pretty_name: Data on Trial — Benchmark Artifacts
---

# Data on Trial — Benchmark Artifacts

Benchmark artifacts for the **data_on_trial** jury-system pipeline. The pipeline code
lives on GitHub: [QuiZet/data_on_trial](https://github.com/QuiZet/data_on_trial).

These files are `.gitignore`d in the code repo because of their size, and mirror the
repository's directory layout so they can be dropped back in place.

## Contents

| Path | Description |
|------|-------------|
| `datasets/` | Downloaded / generated HF datasets used as pipeline inputs (Arrow/JSON). |
| `experiments/` | Experiment snapshots for MMLU-Pro & GPQA: judge results, calibration configs, recovery runs. |
| `analysis/results/` | Aggregated analysis outputs. |
| `analysis/exp/results/` | Logs and outputs from the e1–e13 experiment suite. |
| `analysis/exp/figures/` | Generated figures (PNG). |
| `datagen_related/` | **Archive** — full snapshot of the original production run (the pre-refactor `datagen_related/data_on_trial` working dir): ~1.4 GB of judge results, raw generations, calibration outputs, logs, and SLURM scripts across GPQA & MMLU-Pro. Preserved before the compute cluster was decommissioned. |

## Usage

Clone the code and rehydrate the artifacts alongside it:

```bash
git clone https://github.com/QuiZet/data_on_trial
cd data_on_trial
hf download yungisimon/data_on_trial --repo-type dataset --local-dir .
```

## Notes

No API keys or secrets are included in this dataset. See the GitHub repository for
pipeline code, configuration, and reproduction instructions.