metadata
pretty_name: DedeuceBench (Dev)
license: cc-by-4.0
tags:
- llm-agents
- active-learning
- tool-use
- finite-state-machines
- benchmark
DedeuceBench Dev Split
This dataset provides the public development split for DedeuceBench, an interactive active‑learning benchmark over hidden Mealy machines. Each item in the split is a config entry (seeded episode) defined in levels_dev.json under subset dev.
Summary
- Task: Identification-first — agents must probe a hidden finite-state transducer under a strict query budget using tool calls (
act,submit_table) and submit an exact transition table to succeed. - Deterministic: Episodes are seeded; there is no ground-truth leak in the prompts.
- Split:
devsubset mirrors an easy configuration (e.g., S=2, budget=25) for quick iteration and public sharing.
Files
levels_dev.json: multi-split JSON with adevsubset containing episode items by seed.
Example schema fragment:
{
"splits": {
"dev": {
"mode": "basic",
"trap": false,
"feedback": true,
"budget": 25,
"n_states": 2,
"target_len": 8,
"variety": false,
"items": [ {"seed": 1001}, ... ]
}
}
}
Usage
Programmatic download:
pip install huggingface_hub
python - << 'PY'
from huggingface_hub import hf_hub_download
p = hf_hub_download(
repo_id="comfortably-dumb/dedeucebench-dev",
filename="levels_dev.json",
repo_type="dataset",
)
print(p)
PY
Evaluate with the DedeuceBench CLI (deterministic baseline):
dedeucebench-eval \
--split /path/to/levels_dev.json \
--subset dev \
--model heuristic:none \
--out results.dev.jsonl
dedeucebench-aggregate results.dev.jsonl > leaderboard.dev.csv
Using OpenRouter (OpenAI-compatible):
export OPENAI_API_KEY=... # your OpenRouter key
export OPENAI_BASE_URL=https://openrouter.ai/api/v1
dedeucebench-eval \
--split /path/to/levels_dev.json \
--subset dev \
--provider openrouter \
--model openai/gpt-5-mini \
--out results.openrouter_gpt5mini.dev.jsonl
License
CC-BY-4.0
Citation
- Concept DOI (all versions): 10.5281/zenodo.17166596
- See the main repo README for full citation details.