rfernand/templatic_generation_tasks

Dataset Card for Active/Passive/Logical Transforms

Dataset Summary

This dataset is a synthetic dataset containing a set of templatic generation tasks using both English and random 2-letter words.

Supported Tasks and Leaderboards

[TBD]

Languages

All data is in English or random 2-letter words.

Dataset Structure

The dataset consists of several subsets, or tasks. Each task contains a train split, a dev split, and a test split, and multiple out-of-distribution splits.

Each sample in a split contains a source string, a target string, and an annotation string (describing the sample).

Dataset Subsets (Tasks)

The dataset consists of the following tasks:

- 1_shot_rlw              (1 example input/output pair, a test input, and the gold output, all using random 2-letter words)
- 1_shot_eng              (same as 1_shot_rlw but using English words).
- 1_shot_rlw_10x          (same as 1_shot_rlw, but with 10x the training samples) 
- 2_shot_rlw              (2 example input/output pairs, a test input, and the gold output, all using random 2-letter words)
- 3_shot_rlw              (3 example input/output pairs, a test input, and the gold output, all using random 2-letter words)
- 5_shot_rlw              (5 example input/output pairs, a test input, and the gold output, all using random 2-letter words)
- 10_shot_rtw             (10 example input/output pairs, a test input, and the gold output, all using random 2-letter words)

Data Splits

Most tasks have the following splits:

• train
• dev
• test
• ood_lexical
• ood_cons_count_3
• ood_cons_count_5
• ood_cons_count_7
• ood_cons_count_10
• ood_cons_len_3
• ood_cons_len_5
• ood_cons_len_7
• ood_cons_len_10

Here is a table showing how the number of examples varies by split (for most tasks):

Dataset Split	Number of Instances in Split
train	280,000
dev	35,000
test	35,000
ood_*	84,000

Data Instances

Each sample consits of a source, target, and annotation string (all tab separated).

Here is an example from the train split of the 1_shot_eng task:

{
  'raw': 'Q any mouse ) ; bear A any mouse & . Q road ) ; building A	road & .	{"cons_count": "Q2A1", "cons_len": "Q21.Q11"}'

  'source': 'Q any mouse ) ; bear A any mouse & . Q road ) ; building A',
  'target': 'road & .',
  'annotation': '{"cons_count": "Q2A1", "cons_len": "Q21.Q11"}'
}

Data Fields

• source: the string containing the N-shot examples and the test cue
• target: the string containing the desired (gold) output
• annotation: the string describing the example (as a python or JSON dictionary)

Dataset Creation

Curation Rationale

We wanted a dataset that would test in-context (and from scratch) learning of abstract, semantic-free symbolic transformations, based on a random template for each example. The dataset is designed to test 3 types of out of distribution generalization:

• lexical - known words used in new contexts (relative to train split)
• length - train split uses constituents of 1, 2, or 4 words; OOD splits use 3, 5, 7, or 10 words
• count - train split uses 1, 2, or 4 constituents; OOD splits use 3, 5, 7, or 10 constituents

Source Data

[N/A]

Initial Data Collection and Normalization

[N/A]

Who are the source language producers?

The dataset by generated from templates designed by Paul Smolensky and Roland Fernandez.

Annotations

Besides the source and target strings, each sample contains an annotation string that describes the sample.

Annotation process

The annotation columns were generated from each sample template.

Who are the annotators?

[N/A]

Personal and Sensitive Information

No names or other sensitive information are included in the data.

Considerations for Using the Data

Social Impact of Dataset

The purpose of this dataset is to research how LLM and from-scratch model can learn to solve templatic generation tasks.

Discussion of Biases

[TBD]

Other Known Limitations

[TBD]

Additional Information

The internal name of this dataset is nc_tgt_v11. Also see DATASET_INFO.md and GRAMMAR.md files.

Dataset Curators

The dataset by generated from templates designed by Paul Smolensky and Roland Fernandez.

Licensing Information

This dataset is released under the Permissive 2.0 license.

Citation Information

[TBD]

Contributions

Thanks to The Neurocompositional AI group at Microsoft Research for creating and adding this dataset.