Each request is reviewed by hand. Records keep the license of their source article; about 45% allow commercial use (CC BY, CC BY-SA, CC0) and the rest are non-commercial. The open 1K showcase (Biogod/PMC-ClinicalLogs-1K) needs no approval.

We approve requests that give a real name, an affiliation we can verify and a concrete use (task and model). We decline empty or placeholder entries, and requests for patient care or for redistributing the whole collection. You may reapply with more detail.

PMC Clinical Logs

Time-ordered clinical logs of 150,272 published PubMed Central case reports, each linked to the exact source sentences and the original figures.


Cases	150,272
Events	6,078,136
Original figures	480,791 (72.9 GB)
Source articles	PMC open-access case reports; each record keeps its own license
Clinical validation	None yet: logs are model-generated and not checked by physicians
Reproducible revision	tag `v1.0.1-20261104`

A free 1,000-case sample with images previewable on the Hub: PMC-ClinicalLogs-1K.

What a clinical log is

Each case report is rewritten as an ordered list of events: context, observations, reasoning, actions and results. Every event cites the source text it comes from. Where the article states it, an event also records when it happened and when the information became available. Each event is marked as reported at the time, reconstructed from the text, or a retrospective interpretation. This lets you cut a case at any point and see only what was known by then.

One case, three views. Left: numbered paragraphs of the article (shortened with …). Middle: 9 of the case's 13 events, with links to the paragraphs they cite, the uses links of one reasoning event and the original figure panels; a cut before the MRI hides the later events. Right: the MRI result event as stored. Case: Manjaly et al., "An Unusual Cause of Pseudomedian Nerve Palsy" (2011), PMC3420447, CC BY 3.0; panels are crops of its Fig. 1.

Fields

One row per case in data/cases/*.parquet:

Field	Content
`case_id`, `pmcid`, `title`, `publication_year`, `source_url`	Identity and link to the article
`license_id`, `license_url`, `license_text`	The source article's license and its statement
`clinical_log_json`	The full log as a JSON string (schema `clinical-log/1.4`): `participants`, `events`, `source_notes`
`clinical_log_text`	The same log as readable text
`source_json`	JSON string: article text split into segments, figure captions and metadata (schema `clinical-source/1.0`)
`event_count`, `media_count`, `source_media_count`	Events; figures released; figures in the article
`has_multisubject_events`	The case involves more than one person (for example mother and infant)
`is_showcase`	The case is also in the 1K showcase
`clinical_review_status`	Always `model_generated_not_physician_gold` in this release
`source_draft_sha256`, `source_preparation_sha256`	Checksums linking the row to its build inputs

Main event fields in clinical_log_json: kind (context, observation, reasoning, action, result), phase, text, occurred, available, basis (reported, reconstructed, retrospective), actor, status, links between events (result_of, response_to, uses, after, revises), patient_state.facts, and sources (segment id, character span and the quoted text).

Images: media/*.tar and media-restored/*.tar are TAR shards (each image with a JSON sidecar). media-restored/ holds 68,179 figures released after a figure-parsing correction. data/media/*.parquet has one row per figure (case, caption, license, archive, member, SHA-256) and covers both folders. index/cases.parquet lists every case with its license and shard.

Original figures from CC BY articles in the 1K sample, centre-cropped to squares. Sources, captions and licenses: assets/gallery.json.

Loading

import json
from datasets import load_dataset

cases = load_dataset("Biogod/PMC-ClinicalLogs", "cases", split="train",
                     revision="v1.0.1-20261104", streaming=True, token=True)
case = next(iter(cases))
log = json.loads(case["clinical_log_json"])

# Records whose source license allows commercial use
commercial = cases.filter(lambda r: "nc" not in r["license_id"] and "nd" not in r["license_id"])

Images of one shard, matched to their cases (the shard number is the same across data/cases, data/media and the TAR files):

import tarfile
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

def get(name):
    return hf_hub_download("Biogod/PMC-ClinicalLogs", name, repo_type="dataset", revision="v1.0.1-20261104")

figures = pq.read_table(get("data/media/00074.parquet")).to_pylist()
by_member = {(r["archive"], r["member"]): r for r in figures}
for archive in sorted({r["archive"] for r in figures}):
    with tarfile.open(get(archive)) as tar:
        for m in tar:
            row = by_member.get((archive, m.name))
            if row:
                image_bytes = tar.extractfile(m).read()   # row["case_id"], row["caption"], row["sha256"]

One case with its original images, checksums verified:

hf download Biogod/PMC-ClinicalLogs read_case.py --repo-type dataset --local-dir .
uv run --with huggingface_hub --with pyarrow read_case.py \
  --repo Biogod/PMC-ClinicalLogs --pmcid PMC10001472 --output ./PMC10001472

Licenses

Every record keeps the license of its source article; this collection grants no further rights.

Source license group	Cases	Share
CC BY, CC BY-SA, CC0, public domain (commercial use allowed)	67,417	44.9%
CC BY-NC, CC BY-NC-SA	38,163	25.4%
CC BY-NC-ND	44,473	29.6%
CC BY-ND	219	0.1%

Filter on license_id before any commercial use. Cite this collection and the source articles you use.

Limitations

Not clinically validated. A single language model wrote the logs from each article. Automated checks confirm structure, source quotes and file integrity, not clinical correctness. Event order and times can be wrong, and errors in the original articles are kept.
Excluded content. 1,986 cases whose source license could not be resolved are not released (provenance/pending-source-licenses.json). 1,167 figures with third-party credits are withheld pending rights review (provenance/held-figures.json); their records keep a pointer to the source. An article with no figures is different from a withheld figure.
Multiple subjects. Check cases with has_multisubject_events = true against the source before treating facts as belonging to one patient.
Privacy. The source case reports are already published. This release adds no further de-identification: dates, ages, places and occupations appear as the article states them. Do not try to identify patients.
Overlap with evaluation sets. Some cases are sources of evaluation sets the maintainers are preparing; their identifiers will be listed when those sets are released. The 1K showcase excludes them. If you train on this data, please say so when you report results on benchmarks built from PMC case reports.
Coverage. Case reports over-represent rare and unusual presentations; they are not a sample of routine care.

Release checks and counts: provenance/export-summary.json, provenance/completion.json. Built from the maintainers' snapshot full-20260928-v1.

Citation

@misc{pmc_clinical_logs_2026,
  title  = {PMC Clinical Logs},
  author = {Ji, Jiahao and Yu, Jintai and Yuan, Zhiyuan},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Biogod/PMC-ClinicalLogs}},
  note   = {Revision v1.0.1-20261104}
}

Downloads last month: 22,055

Total file size:

77.1 GB