Datasets:

aieng-lab
/

gradiend-function-composition

license: cc-by-4.0
configs:
  - config_name: default
    data_files:
      - split: train
        path: default/train.parquet
      - split: validation
        path: default/validation.parquet
      - split: test
        path: default/test.parquet
    default: true
  - config_name: medium
    data_files:
      - split: train
        path: medium/train.parquet
      - split: validation
        path: medium/validation.parquet
      - split: test
        path: medium/test.parquet
  - config_name: hard
    data_files:
      - split: train
        path: hard/train.parquet
      - split: validation
        path: hard/validation.parquet
      - split: test
        path: hard/test.parquet

GRADIEND Function Composition Data

Synthetic alias-resolution cloze data used in the GRADIEND/ACTIEND/SAE/CAA comparison (aieng-lab/iend-study).

Usage

from datasets import load_dataset

ds = load_dataset("aieng-lab/gradiend-function-composition", "default", split="train")

Splits: train, validation, test.

Dataset Details

Dataset Description

One variable aliases a variable holding the target value, e.g. m = fork; p = seal; h = m; h = [MASK] (target fork). The factual class is RESULT.

Dataset Structure

Configs default (easy), medium, and hard; each has all splits (8,000 / 1,000 / 1,000 rows).

masked: the input text with a single [MASK] slot to be predicted
label: the factual target token(s) for the slot
label_class: class of the factual target
alternative: the alternative (counterfactual) target for the slot
alternative_class: class of the alternative target
feature_class_id: identifier of the feature class (equal to label_class)
split: train, validation, or test
difficulty: difficulty level of the row

default = easy (one alias step, canonical format). medium (chain depth 2) and hard (chain depth 3) add longer alias chains and format variation.

Dataset Sources

Repository: https://github.com/aieng-lab/iend-study
Generated with study.data.synthetic in the repository above; no external data.

Dataset Creation

Variable names are sampled without replacement from the 26 lowercase Latin letters; values come from the same split-disjoint 75-value vocabulary as gradiend-key-value. Easy examples contain one alias step and zero or one unrelated distractor assignment. If a distractor is present its value is the alternative, otherwise a different value from the split-local vocabulary is used. Alias and distractor variables are distinct, distractor values never equal the target, and duplicate masked contexts are removed.

Bias, Risks, and Limitations

Fully synthetic and templated; it probes one narrow symbolic mechanism, not general reasoning, and results need not transfer to natural text. Value/token vocabularies are small and hand-curated.

Citation

BibTeX: TODO