jessicalamjh/discussion-generation

PMCOA Discussion Generation Dataset

A dataset of 627 biomedical papers from PubMed Central Open Access, built for the task of discussion section generation: given a manuscript (with its Discussion section removed) and the full text of its cited papers, generate the Discussion section.

Each sample contains:

• manuscript — the paper with its Discussion section removed
• relevant_papers — full text of the papers cited in the gold discussion
• gold_discussion — the ground-truth Discussion section

The canonical schema is defined in src/discussion_generation/data/schemas.py and documented in detail in src/discussion_generation/data/README.md.

---

Why some fields are JSON strings

Apache Arrow (which backs HuggingFace datasets) requires every column to have a fixed, uniform schema. Two patterns in the native schema are incompatible with that:

Pattern	Arrow's problem
dict[PaperIdType, ...] — keys are an open str enum	Arrow infers column types from the first record; unseen key names in later records break the schema
list[Content] where Content = Paragraph | Section — recursive, polymorphic	Arrow cannot represent recursive or union-typed nested structs

The following fields are serialized to JSON strings before upload:

HF column	Native type	Reason
manuscript	Paper	recursive / polymorphic, plus dynamic dict keys in nested all_known_paper_ids/bibliography
relevant_papers	dict[BibliographyEntryId, Paper]	dynamic dict keys at top level
gold_discussion	Section	recursive / polymorphic

Additionally, ContentId (tuple[int, ...]) is stored as list[int] because Arrow has no tuple type. This affects gold_discussion_content_id and each element of content_ids_referenced_in_gold_discussion.

All other fields keep their original structure.

---

Restoring the original structure

Parse JSON strings back and validate with the Sample Pydantic model. Pydantic handles list[int] → tuple[int, ...] coercion for ContentId fields automatically.

import json
from datasets import load_dataset
from discussion_generation.data.schemas import Sample

ds = load_dataset("jessicalamjh/discussion-generation", split="train")

def restore(record: dict) -> Sample:
    record = dict(record)

    record["manuscript"]      = json.loads(record["manuscript"])
    record["relevant_papers"] = json.loads(record["relevant_papers"])
    record["gold_discussion"] = json.loads(record["gold_discussion"])

    return Sample.model_validate(record)

samples: list[Sample] = [restore(r) for r in ds]

Or, more simply, via load_schematized_dataset in src/discussion_generation/utils/data.py, which does the same thing:

from discussion_generation.utils.data import load_schematized_dataset

samples = load_schematized_dataset("jessicalamjh/discussion-generation", split="train")