klusai/open-weight-judges-eval
Open-Weight Judges Evaluation Dataset Evaluation outputs from a study on cross-task transferability of family-diverse open-weight judge panels for synthetic text evaluation. Overview This dataset contains ~6,180 judge evaluations across three synthetic-text tasks, produced by a three-member open-weight panel, proprietary baselines (GPT-o4-mini, Gemini 2.5 Flash), a specialized open evaluator (Atla Selene Mini), and four bias audits. Tasks Task… See the full description on the dataset page: https://huggingface.co/datasets/klusai/open-weight-judges-eval.
Open-Weight Judges Evaluation Dataset
Evaluation outputs from a study on cross-task transferability of family-diverse open-weight judge panels for synthetic text evaluation.
Overview
This dataset contains ~6,180 judge evaluations across three synthetic-text tasks, produced by a three-member open-weight panel, proprietary baselines (GPT-o4-mini, Gemini 2.5 Flash), a specialized open evaluator (Atla Selene Mini), and four bias audits.
Tasks
Panel Members
Panel members are deliberately from families disjoint from all generator models.
Directory Structure
tf1_panel/ # Panel evaluations: 3 judges x 1000 items
tf2_panel/ # Panel evaluations: 3 judges x 96 items
tf3_panel/ # Panel evaluations: 3 judges x 100 items
tf1_proprietary/ # GPT-o4-mini on 1000 TF1 items
tf2_proprietary/ # GPT-o4-mini on 96 TF2 items
tf3_proprietary/ # GPT-o4-mini + Gemini Flash on 100 TF3 items each
tf1_selene/ # Selene Mini on 200 TF1 items
tf2_selene/ # Selene Mini on 96 TF2 items
tf3_selene/ # Selene Mini on 100 TF3 items
bias_audit/ # Position, stability, length, family audit results
paper_tables/ # Pre-computed analysis tables (CSV)
tf1_samples.jsonl # Input fables (TF1)
tf2_samples.jsonl # Input translation pairs (TF2)
tf3_samples.jsonl # Input Romanian fables (TF3)Record Format
Panel and proprietary outputs use JSONL with one record per item:
{
"item_id": "...",
"judge_model": "granite4.1:30b",
"task": "tf1_generation",
"scores": {"grammar": 8, "creativity": 6, ...},
"input_tokens": 1714,
"output_tokens": 164,
"latency_ms": 27370.9,
"timestamp": "2026-05-01T21:28:39"
}TF2/TF3 scores use nested format: {"accuracy": {"score": 5, "justification": "..."}}.
Key Findings
- System-level rank correlation between panel and GPT-o4-mini: rho=0.927 (p<0.001) on TF1's 10 generators
- Item-level Krippendorff alpha near zero across all tasks (score compression)
- EXAONE 3.5 most position-stable (100% grammar match) and run-stable (91-100%)
- Granite 4.1 shows significant length bias (rho=-0.27 for grammar)
- Zero family-preference bias by construction (disjoint panel-generator families)
Citation
@article{nadas2026openjudges,
title={Cross-Task Transferability of Family-Diverse Open-Weight Judge Panels for Synthetic Text Evaluation},
author={Nada\c{s}, Mihai and Dio\c{s}an, Laura},
year={2026},
note={In preparation}
}Related
- Code: klusai/open-weight-judges
- TF1 dataset: klusai/ds-tf1-en-3m
- TF2 dataset: klusai/ds-tf2-en-ro-3m
- TF3 dataset: klusai/ds-tf3-ro-3m
