README.md

metadata

license: mit
task_categories:
  - reinforcement-learning
tags:
  - eleusis
  - verifiers
  - tool-use
  - rule-induction
dataset_info:
  features:
    - name: rule_id
      dtype: string
    - name: label
      dtype: string
    - name: family
      dtype: string
    - name: code
      dtype: string
    - name: split
      dtype: string
    - name: uses_mainline
      dtype: bool
    - name: accepted_cards
      list: string
    - name: rejected_cards
      list: string
    - name: empty_mainline_accepted_cards
      list: string
    - name: empty_mainline_rejected_cards
      list: string
    - name: representative_acceptance_rate
      dtype: float64
    - name: rule_index
      dtype: int64
    - name: split_index
      dtype: int64
    - name: dataset_version
      dtype: string
    - name: source_split
      dtype: string
    - name: source_dataset_version
      dtype: string
    - name: oracle_max_turns
      list: int64
    - name: oracle_optimal_turns
      list:
        list: float64
    - name: oracle_expected_optimal_turns
      list:
        list: float64
    - name: oracle_methods
      list:
        list: string
  splits:
    - name: train
      num_bytes: 81960
      num_examples: 66
    - name: eval
      num_bytes: 32356
      num_examples: 26
  download_size: 116311
  dataset_size: 114316
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: eval
        path: data/eval-*

Eleusis HF Rules

This dataset contains the HF-like Eleusis rule curriculum previously embedded in nph4rd/eleusis as v8-relational-curriculum-90, republished as a standalone rule dataset so the environment can switch rule sets by dataset repo.

Dataset version: v1-hf-like-train-hf-eval-64

Splits

• train: 38 HF-like curriculum rules.
• eval: 26 HF benchmark-style rules, copied from the old internal hf split.

The dataset intentionally exposes only train and eval. A training run that uses this dataset and evaluates on split=eval is therefore evaluating directly on the HF benchmark-style rules.

Family Distribution

Train:

• pair_position: 2
• previous_card: 23
• static_combo: 3
• static_rank: 6
• static_suit_color: 4

Eval:

• pair_position: 2
• previous_card: 11
• static_combo: 3
• static_rank: 6
• static_suit_color: 4

Schema

Each row contains:

• rule_id: stable identifier.
• label: human-readable description for inspection only.
• family: rule family.
• code: Python predicate body or expression used by the environment verifier.
• split: published split (train or eval).
• source_split: original split from the embedded curriculum.
• dataset_version: this dataset version.
• accepted_cards / rejected_cards: empty-mainline card partitions.
• representative_acceptance_rate: acceptance rate over representative mainlines.

The environment should use code for scoring and should not expose label or code to the model.