oroikono/Kagamaya-PDEBench

Kagamaya: reproduce and compare evidence-selection rules

Paper · Code · Citation

How does choosing a different numerical summary change an equation verifier's answer? Use this dataset to reproduce the paper's PDEBench comparison, inspect the errors, or evaluate your own rule for choosing among the same twelve summaries.

1. Reproduce the published comparison

pip install huggingface_hub==2.4.0
hf download oroikono/Kagamaya-PDEBench --type dataset \
  --revision v1.1.0 --local-dir Kagamaya-PDEBench
cd Kagamaya-PDEBench
python3 reproduce.py

After download, the check uses only Python's standard library. No GPU, API key or model calls are needed. It checks the stored numerical records and reports status: PASS with these counts:

Selection rule	Correct	Wrong identification	Wrong rejection	No verdict
Maximize the public score	828	69	3	0
Always use the fixed fine view	900	0	0	0
Require four views to agree	645	0	0	255

All rules are evaluated on the same 900 episodes. Cases with no verdict remain in the denominator.

2. Try your own selection rule

Start with a runnable example:

python3 example_policy.py --policy fixed_fine > ../my_choices.jsonl
python3 evaluate.py --choices ../my_choices.jsonl

Your rule can instead produce one JSON object per episode, with exactly episode_id and view_id. Use one of that episode's twelve view IDs, or null to withhold a verdict. Submit all 900 episodes; missing, duplicate or invalid choices are rejected. Keep the choices file outside the downloaded dataset folder.

The evaluator reports correct answers, both kinds of error, abstentions and coverage. It uses the existing verifier answers; it does not fit new equations. The example also supports score_max and all_mode_agreement.

3. Load the tables in Python

pip install datasets==5.1.0 pyarrow==26.0.1

from datasets import load_dataset

views = load_dataset("oroikono/Kagamaya-PDEBench", "views",
                     split="evaluation", revision="v1.1.0")

Configuration	Rows	Contents
views	10,800	Twelve numerical summaries and verifier answers per episode
episodes	900	Episode information and the separate numerical reference
decisions	2,700	The three published rules' choices and outcomes

Join on episode_id. The 900 episodes come from 60 source trajectories, not 900 independent trajectories. Labels are public: this is a reusable, retrospective comparison, not a hidden test set.

Which paper results can I reproduce here?

This download covers PDEBench v5 only. The main 1-D benchmark, Claude/Qwen experiments and other controls have separate materials on GitHub. Paper-to-artifact guide · Data details and provenance

These commands replay stored results. They do not regenerate PDE trajectories, rerun language models, or establish performance on new systems.

License and citation

Tables: CC BY 4.0, derived from PDEBench Datasets; keep the attribution in NOTICE.md. Author-owned code and documentation: MIT.

@misc{oikonomou2026verifier,
  title = {{The Verifier Is Part of the Simulator: Evidence-View Gaming in Agentic Equation Discovery}},
  author = {Oikonomou, Orestis and Raonic, Bogdan},
  year = {2026},
  eprint = {2611.01334},
  archivePrefix = {arXiv},
  primaryClass = {cs.LG},
  url = {https://arxiv.org/abs/2611.01334}
}