laya-marker-corpus

Training corpus for a typed-decision head: the model is never asked to generate text, only to score a fixed set of options handed to it together with the question.

The point of the mixture is breadth, not any single task. A head trained on one task learns that task; the aim here is a head that learns to read the instruction, so it is trained on 368 of them at once and measured on tasks it has never seen.

Files

file	rows	tasks	types
mix_big.jsonl	72,537	174	choice 35,844 · noul 35,851 · multi 842
mix_tables.jsonl	23,280	194	choice 11,640 · noul 11,640
total	95,817	368

Row schema

{
  "task": "bb_implicit_relations",
  "state": "…the context the question is asked about…",
  "instruction": "Task: implicit relations. Choose the correct option.",
  "type": "choice",
  "options": ["Brother", "Sister", "Colleague", "…"],
  "answer": "Sister"
}

• state — the context. For tables it is {"table": name, "record": {…}} with one field masked.
• instruction — what to decide, in natural language, not from a fixed vocabulary.
• type — choice (one of K), noul (yes/no), multi (any subset; answer is a list).
• options — the full candidate set. Option text is meant to live in the context, so structured and negated options ("billing: charges, invoices. NOT order tracking") survive.
• answer — a member of options, or a list of members when type is multi.

A fourth type, score (rate on a scale), is not present in this build.

Provenance and licensing

This corpus is a derivative of two upstream sources with different terms. There is no single license covering all of it; check the upstream dataset before redistributing a subset.

mix_big.jsonl — derived from BIG-bench multiple-choice tasks, Apache 2.0.

mix_tables.jsonl — derived from the mstz/* re-uploads of UCI-style tabular datasets. One column per record is masked and becomes the question; a column qualifies when its distinct-value count is within [min_k, max_k] and below 30% of the row count. The 19 tables:

abalone  adult  bank  breast  car  compas  german  glass  haberman  heart_failure
heloc  ionosphere  page_blocks  pima  post_operative  spambase  speeddating
student_performance  wine

17 of these carry a Creative Commons tag upstream; bank and speeddating are tagged unknown. Attribution belongs to the original UCI contributors in every case.

A note on the tabular sources

Several are standard fairness-research datasets and contain sensitive attributes: compas (criminal recidivism), adult (census income), german (credit scoring) and speeddating (personal and demographic attributes). They are included here only as a source of structured masked-cell prediction problems — a way to make the head read a schema and a record. They are well documented as carrying historical and demographic bias, and nothing here corrects for it. Do not treat a model trained on this corpus as fit for decisions about people.

Held-out sets

clinc151, mmlu, qnli and yelp5 are deliberately absent and exist only to measure transfer to unseen tasks. qnli and yelp5 are sealed further: opened once, at the end, and never used to pick a configuration.