Templatic Generation Tasks dataset (Jan-08-2025)

A synthetic dataset for containing examples of the Templatic Generation Task, as described in our Oct 2024 Technical report: Mechanisms of Symbol Processing for In-context Learning in Transformer Networks.

Each example has 2-3 parts, separated by TAB character: x - the example input y - the example label info - OPTIONAL text identifying the type of example (for optional filtering during training)

Currently available tasks:

1_shot_rlw:          each example is 1 sample input/output pair + input cue; random two-letter words
2_shot_rlw           each example is 2 sample input/output pairs + input cue; random two-letter words
3_shot_rlw:          each example is 3 sample input/output pairs + input cue; random two-letter words
5_shot_rlw:          each example is 5 sample input/output pairs + input cue; random two-letter words
10_shot_rlw:         each example is 10 sample input/output pairs + input cue; random two-letter words

1_shot_eng:          each example is 1 sample input/output pair + input cue; english words
1_shot_rlw_10x:      same as 1_shot_rlw, but with 10x as many training examples

Each task contains the following splits:

train               contains 1, 2, or 4 constituents; each with 1, 2, or 4 parts
dev                 contains 1, 2, or 4 constituents; each with 1, 2, or 4 parts
test                contains 1, 2, or 4 constituents; each with 1, 2, or 4 parts
ood_lexical         the constituent part vocab is held out from training (except for the echo examples)
ood_cons_len_3      all templates have constituents have 3 parts
ood_cons_len_5      all templates have constituents have 5 parts
ood_cons_len_7      all templates have constituents have 7 parts
ood_cons_len_10     all templates have constituents have 10 parts
ood_cons_count_3    all templates have 3 constituents
ood_cons_count_5    all templates have 5 constituents
ood_cons_count_7    all templates have 7 constituents
ood_cons_count_10   all templates have 10 constituents

Definitions: - echo examples: use to introduce out-of-distribution vocabulary symbols to the model (in the train split) - template - used to generate an example (input/output sample pairs, cue input, gold output)

Normal example: Q oy xf kq be ? jp A jp = . Q jf ty zu np ? cx A cx = . {"cons_count": "Q2A1", "cons_len": "Q41.Q41"}

breakdown:
    1 sample input:     Q oy xf kq be ` ? jp A
    1 sample output:    jp = .
    cue input:          Q jf ty zu np ` ? cx A
    gold output:        cx = .
    example info:       {"cons_count": "Q2A1", "cons_len": "Q41.Q41"}

Echo example: Q ZW A ZW . Q VI A VI . {"type": "echo"}

breakdown:
    1 sample input:     Q ZW A
    1 sample output:    ZW .
    cue input:          Q VI A
    gold output:        VI .
    example info:       {"type": "echo"}