README.md

metadata

license: mit
tags:
  - datasets
  - evaluation
  - reinforcement-learning
  - tool-calling
  - synthetic
  - verifiers
  - agents
  - benchmark
configs:
  - config_name: mealy_n4_b100
    data_files:
      - split: test
        path: mealy/hf_mealy_n4_b100_test.jsonl
      - split: train
        path: mealy/hf_mealy_n4_b100_train.jsonl
  - config_name: protocol_n4_e5_b120
    data_files:
      - split: test
        path: protocol/hf_protocol_n4_e5_b120_test.jsonl
      - split: train
        path: protocol/hf_protocol_n4_e5_b120_train.jsonl
  - config_name: apienv_n7_e7_b200
    data_files:
      - split: test
        path: apienv/hf_apienv_n7_e7_b200_test.jsonl
      - split: train
        path: apienv/hf_apienv_n7_e7_b200_train.jsonl
  - config_name: exprpolicy_b60
    data_files:
      - split: test
        path: exprpolicy/hf_exprpolicy_b60_test.jsonl
      - split: train
        path: exprpolicy/hf_exprpolicy_b60_train.jsonl

DedeuceRL Splits

Task splits for the DedeuceRL benchmark — a modular framework for active system identification and code debugging tasks.

• Source: https://github.com/AashVed/DedeuceRL
• License: MIT

Skins

Skin	Task	Description
Mealy	System Identification	Identify hidden Mealy machine (finite-state transducer)
Protocol	API Reverse Engineering	Discover state-dependent REST API behavior
APIEnv	SaaS API Identification	Realistic stateful SaaS API with variants
ExprPolicy	Code Debugging	Fix buggy typed policy DSL expression

Splits Summary

Skin	Config	Test	Train
Mealy	n_states=4, budget=100	100 (0-99)	10,000 (100-10099)
Protocol	n_endpoints=5, n_states=4, budget=120	100 (0-99)	10,000 (100-10099)
APIEnv	n_endpoints=7, n_states=7, budget=200	100 (0-99)	10,000 (100-10099)
ExprPolicy	budget=60, n_public=8, n_hidden=80	100 (0-99)	10,000 (100-10099)

All splits generated with trap=False.

Usage

Load with HuggingFace Datasets

from datasets import load_dataset

# Load Mealy training set
ds = load_dataset("comfortably-dumb/DedeuceRL", "mealy_n4_b100", split="train")

# Load ExprPolicy test set
ds_test = load_dataset("comfortably-dumb/DedeuceRL", "exprpolicy_b60", split="test")

Run with DedeuceRL CLI

dedeucerl-eval \
  --skin mealy \
  --split mealy/dedeucerl_mealy_n4_b100_test.json \
  --model openai:gpt-4o \
  --out results.jsonl

Schema

Each episode contains:

• skin: Environment type
• subset: Split name (train, test)
• seed: Reproducibility seed
• prompt: System + user messages for the agent
• answer: JSON-encoded ground truth

Generation Parameters

Skin	Parameters
Mealy	n_states=4, budget=100, trap=false
Protocol	n_endpoints=5, n_states=4, budget=120, trap=false
APIEnv	n_endpoints=7, n_states=7, budget=200, trap=false
ExprPolicy	budget=60, n_public=8, n_hidden=80, trap=false

Citation

@software{dedeucerl2025,
  title = {DedeuceRL: A Modular Framework for Active System Identification Benchmarks},
  author = {Vedansh},
  year = {2025},
  url = {https://github.com/AashVed/DedeuceRL},
  doi = {10.5281/zenodo.18280315}
}