Datasets:
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is a compilation of ClinicalTrials.gov data (a U.S. Government
database) by the dataset author. The compilation is licensed under
CC-BY-NC-4.0. As a condition of accessing these files, you agree to:
- Cite this dataset, and attribute the underlying data to ClinicalTrials.gov,
in any publication or derivative work. - Not use these files, or copies of them, for commercial purposes without a
separate commercial license from the author. Examples: ML training data
for commercial models, redistribution in a commercial product or service,
inclusion in databases offered for sale, or quantitative trading /
commercial market analysis.
These terms apply to this dataset only, not to data you obtain directly from
ClinicalTrials.gov. For commercial licensing, contact licensing@anamnesisdata.com
or visit https://anamnesisdata.com.
Log in or Sign Up to review the conditions and access this dataset content.
This dataset has moved to anamnesis-data/clinical_trials_history. New releases are published there only; this repository is no longer updated. Access approvals do not carry over, so please request access on the new repository. Existing tags here (up to
v2026.09.26) remain available for pinned use. The data files were removed from the main branch on 2026-11-03; they have defects fixed in the new repository.
ClinicalTrials.gov Version History
A longitudinal, per-version snapshot of every protocol amendment ever submitted to ClinicalTrials.gov, normalized into a relational schema. Unlike most ClinicalTrials.gov mirrors that publish only the latest snapshot of each trial, this dataset preserves every version of every trial — making it suitable for event studies, change-tracking, and any analysis where the dynamics of protocol amendments matter.
Published by Anamnesis Data. Free for academic and non-commercial research; commercial licenses are available (see License).
What's in the core config
The core config is one row per (nct_id, nct_version) containing the scalar
fields from each trial's protocol-section snapshot. 4,444,542 rows covering
~605K trials across their full version histories from 1999 to today.
from datasets import load_dataset
ds = load_dataset("brbk/clinical_trials_history", "core")
print(ds["train"][0])
# {'nct_id': 'NCT00000125', 'nct_version': 0, 'overall_status': 'ACTIVE_NOT_RECRUITING', ...}
The config has 96 protocol-section scalar columns. The full column list is in
the parquet schema, viewable via the HuggingFace Datasets Viewer or
ds.features.
Key columns
Trial identification and version:
| Column | Type | Description |
|---|---|---|
nct_id |
string | ClinicalTrials.gov identifier |
nct_version |
int64 | Per-trial version number, aligned 1:1 with CT.gov's 0-based version index. Sorting by (nct_id, nct_version) walks each trial's history chronologically. |
ct_gov_version |
int64 | Equals nct_version. Retained as a separate column for backward compatibility with consumers that switched on it during earlier schema versions. |
Status and design:
| Column | Type | Description |
|---|---|---|
overall_status |
string | Recruitment / completion status |
why_stopped |
string | Reason given for stopping, if stopped |
study_type |
string | INTERVENTIONAL / OBSERVATIONAL / EXPANDED_ACCESS |
allocation |
string | Randomized vs non-randomized |
intervention_model |
string | Parallel, crossover, single-group, etc. |
primary_purpose |
string | Treatment, prevention, diagnostic, etc. |
masking |
string | Blinding level |
enrollment_count |
int64 | Enrollment (anticipated or actual) |
enrollment_type |
string | ACTUAL vs ESTIMATED |
has_results |
bool | TRUE if this version has posted results |
Sponsorship:
| Column | Type | Description |
|---|---|---|
lead_sponsor_name |
string | Lead sponsor |
lead_sponsor_class |
string | INDUSTRY / FED / NIH / OTHER_GOV / NETWORK / AMBIG / OTHER / INDIV / UNKNOWN |
responsible_party_type |
string | Sponsor, principal investigator, etc. |
organization_name |
string | Submitting organization |
organization_class |
string | Submitting organization's class |
Dates (read the Date Field Semantics section below):
| Column | Type | Description |
|---|---|---|
start_date |
date | Trial start; start_date_precision indicates day/month/year |
completion_date |
date | Overall completion |
primary_completion_date |
date | Primary endpoint completion |
study_first_submit_date |
date | First-ever submission to CT.gov (constant across versions) |
study_first_post_date |
date | First-ever public posting (constant across versions) |
last_update_submit_date |
date | Date THIS version was submitted by registrant |
last_update_post_date |
date | Date THIS version was posted publicly. Use this for event-study / public-visibility timing. |
last_update_post_date_type |
string | ACTUAL (precise) vs ESTIMATED (approximate, common pre-2018) |
The remaining ~70 columns cover titles, brief/detailed descriptions,
eligibility criteria, age ranges, design-info detail, results-disclosure
dates, IPD-sharing fields, and other protocol-section scalars. See
ds.features for the full list with parquet-native types.
How versions work
Every time a registrant amends a trial's protocol or status on
ClinicalTrials.gov, CT.gov assigns a new version. This dataset captures each
chronological version as its own row keyed on (nct_id, nct_version). The
nct_version index is 0-based and aligns directly with CT.gov's own version
numbering — nct_version=0 is the trial's first submission, and the highest
nct_version per trial is the latest amendment we've captured.
For each trial, sorting by nct_version produces the chronological version
sequence. No special filtering is needed.
Date Field Semantics
This is the most important thing to know about the dataset:
last_update_post_dateis per-version, not stale. Each version's row carries the post date for that version, not a value carried forward from a prior version.last_update_submit_date≠last_update_post_date. Submit is when the registrant uploaded the change; post is when CT.gov published it. The gap ranges from 1 day (auto-estimated) to ~30 days (review queue). For event studies tied to market reactions,last_update_post_dateis what the public could see.last_update_post_date_typematters. Older versions (pre-~2018) carryESTIMATEDpost dates that are mechanical (often submit+1 day). Newer versions areACTUAL. TreatESTIMATEDas approximate when timing precision matters.
| Question | Field |
|---|---|
| When did the registrant submit this version? | last_update_submit_date |
| When did the public first see this version? | last_update_post_date |
| How precise is that post date? | last_update_post_date_type |
| When was the trial first registered? | study_first_post_date |
What's in the interventions config
One row per (nct_id, nct_version, intervention_name, intervention_type) —
the registrant-entered intervention list for each trial-version. 9,401,447
rows covering ~544K trials across 3.8M trial-versions. Trials with no
interventions (most observational studies) simply have no rows here.
from datasets import load_dataset
interventions = load_dataset("brbk/clinical_trials_history", "interventions",
split="train")
print(interventions[0])
# {'nct_id': 'NCT00000125', 'nct_version': 0,
# 'intervention_name': '...', 'intervention_type': 'DRUG', 'description': '...'}
Columns
| Column | Type | Description |
|---|---|---|
nct_id |
string | Joins to core.nct_id |
nct_version |
int64 | Joins to core.nct_version |
intervention_name |
string | Registrant-entered name (free text — not normalized to a vocabulary) |
intervention_type |
string | One of 11 enum values; see breakdown below |
description |
string | Free-text description of the intervention; ~80% of rows have one |
intervention_type values
Distribution across all rows:
| Type | Count | % |
|---|---|---|
| DRUG | 4,950,405 | 52.7% |
| PROCEDURE | 1,223,174 | 13.0% |
| OTHER | 1,093,004 | 11.6% |
| BIOLOGICAL | 656,714 | 7.0% |
| DEVICE | 530,594 | 5.6% |
| BEHAVIORAL | 491,012 | 5.2% |
| RADIATION | 170,096 | 1.8% |
| DIETARY_SUPPLEMENT | 120,653 | 1.3% |
| DIAGNOSTIC_TEST | 96,111 | 1.0% |
| GENETIC | 47,608 | 0.5% |
| COMBINATION_PRODUCT | 22,076 | 0.2% |
Joining to core
Each interventions row joins back to core on (nct_id, nct_version) —
so you can attach trial-level scalars (status, dates, sponsor, phase) to any
intervention slice. Example: drug interventions on completed Phase 3 trials.
import polars as pl
from datasets import load_dataset
core = pl.from_arrow(load_dataset(
"brbk/clinical_trials_history", "core", split="train"
).data.table)
ix = pl.from_arrow(load_dataset(
"brbk/clinical_trials_history", "interventions", split="train"
).data.table)
drug_phase3 = (
ix.filter(pl.col("intervention_type") == "DRUG")
.join(core.select("nct_id", "nct_version", "overall_status",
"primary_completion_date"),
on=["nct_id", "nct_version"])
.filter(pl.col("overall_status") == "COMPLETED")
)
Caveat —
intervention_nameis free text. Two rows with names"Pembrolizumab","KEYTRUDA", and"MK-3475"all refer to the same drug. Resolving names to a controlled vocabulary (UNII / RxCUI / ChEMBL ID) requires an entity-resolution step that is not in this dataset.
Why this exists
Existing public ClinicalTrials.gov datasets — including AACT, the OpenFDA mirror, and HuggingFace's flat snapshot mirrors — publish only the current state of each trial. Anything that requires historical version data (event studies on protocol amendments, longitudinal modeling of trial trajectories, change-frequency analyses) has had no ready-made source. This dataset publishes that history directly.
Example queries
Find all trials whose status changed within the last 30 days
import polars as pl
from datasets import load_dataset
ds = load_dataset("brbk/clinical_trials_history", "core", split="train")
df = pl.from_arrow(ds.data.table)
changes = (
df.sort(["nct_id", "nct_version"])
.with_columns(prev_status=pl.col("overall_status").shift().over("nct_id"))
.filter(
pl.col("overall_status") != pl.col("prev_status"),
pl.col("prev_status").is_not_null(),
pl.col("last_update_post_date") >= pl.date(2026, 4, 8),
)
.select("nct_id", "nct_version", "prev_status", "overall_status",
"last_update_post_date")
)
Completion-date pushes ("trial slipped")
slips = (
df.sort(["nct_id", "nct_version"])
.with_columns(prev_completion=pl.col("completion_date").shift().over("nct_id"))
.filter(pl.col("completion_date") > pl.col("prev_completion"))
.with_columns(slip_days=(pl.col("completion_date") - pl.col("prev_completion")).dt.total_days())
.select("nct_id", "nct_version", "prev_completion", "completion_date",
"slip_days", "last_update_post_date")
)
Update cadence
Refreshed periodically. Each refresh is committed to a Git tag
(v2026.09.26, ...) so consumers can pin to a specific snapshot:
ds = load_dataset("brbk/clinical_trials_history", "core",
revision="v2026.09.26")
The most recent commit on main is always the latest snapshot.
Changelog
v2026.09.26
core: 4,444,542 rows covering 604,583 trials, reflecting the ClinicalTrials.gov release of 2026-09-25.- New
interventionsconfig: 9,401,447 rows covering ~544K trials. - Earlier revisions (
v2026.05.08,v2026.05.15) have been removed from this repository and can no longer be pinned.
Planned future configs
Same dataset URL, joined either on nct_id (per-trial configs) or
(nct_id, nct_version) (per-version configs) — these will be added as
additional configs without disrupting core:
| Config | Grain | Join key |
|---|---|---|
history_metadata |
One row per trial: trial-level history metadata | nct_id |
version_history |
One row per version: change date and status | nct_id, nct_version |
history_original_outcomes |
Outcome measures as first submitted | nct_id |
version_changes |
Field-level changes between consecutive versions | nct_id, nct_version |
outcomes |
Primary/secondary outcomes per version | nct_id, nct_version |
eligibility |
Per-version eligibility text and age/sex fields | nct_id, nct_version |
locations |
Per-version trial site locations | nct_id, nct_version |
sponsors |
Per-version lead sponsor and collaborators | nct_id, nct_version |
results |
Outcome measurements and adverse events for trials with results | varies |
The core config will not break — additive changes only (new columns may
appear; no renames or type changes within the published schema). Breaking
changes will be announced and gated behind a major version bump.
Source
The trial records in this dataset come from ClinicalTrials.gov, a database operated by the U.S. National Library of Medicine (NLM) at the National Institutes of Health.
Data currency: this release reflects ClinicalTrials.gov data as processed by ClinicalTrials.gov on 2026-09-25. ClinicalTrials.gov is updated daily; check ClinicalTrials.gov for current information.
Modifications made to the ClinicalTrials.gov data
- Restructured: records were reshaped into tables (
core: one row per(nct_id, nct_version);interventions: one row per(nct_id, nct_version, intervention_name, intervention_type)). - Subset of fields: each config publishes a subset of the ClinicalTrials.gov fields; the Parquet schema of each config lists exactly which.
- Version index:
nct_versionequals ClinicalTrials.gov's own 0-based version number. - Partial dates: dates reported only as
YYYY-MMorYYYYare stored as the first day of that period (YYYY-MM-01,YYYY-01-01); where a*_precisioncolumn exists, it records the original precision. Date values in any other format are stored as null. - Interventions: entries missing a name or a type are omitted, and duplicate
(intervention_name, intervention_type)entries within one version are collapsed to the first occurrence (the later entries' descriptions are not kept). - Always-null columns: in
core,expanded_access_status_for_nct_id,version_holder,first_mcp_post_date,first_mcp_post_date_typeandestimated_results_first_submit_dateare null on every row.
Apart from the changes listed above, field values are carried over as returned by ClinicalTrials.gov.
License
The underlying ClinicalTrials.gov records
The trial records themselves (registration and results information submitted by study sponsors) are from ClinicalTrials.gov, a United States Government database. No proprietary rights are claimed in them. They are available to anyone at no charge directly from ClinicalTrials.gov, subject to the ClinicalTrials.gov Terms and Conditions, which note that some ClinicalTrials.gov data may be subject to the copyright of third parties. Nothing on this page restricts your use of data you obtain from ClinicalTrials.gov.
This dataset
The compilation published here (its relational structure, schema, version history, derived fields, and documentation) is licensed under CC BY-NC 4.0 to the extent the author holds rights in it.
Access terms
Access to the files in this repository is gated. As a condition of access, you agree to:
- Cite this dataset (see Citation) in any publication or derivative work that uses it, and attribute the underlying data to ClinicalTrials.gov.
- Not use the files distributed here, or copies of them, commercially without a separate commercial license. Examples: redistributing them in a commercial product or service, using them as training data for commercial machine-learning models, including them in databases offered for sale, or using them for quantitative trading or commercial market analysis.
These access terms are an agreement between you and the author about this dataset. They are separate from the CC BY-NC 4.0 license and from the ClinicalTrials.gov Terms and Conditions, and they do not apply to data you obtain directly from ClinicalTrials.gov.
For commercial licensing, contact licensing@anamnesisdata.com or visit anamnesisdata.com.
Citation
If you use this dataset in research, please cite:
@misc{clinicaltrials_history_2026,
title = {ClinicalTrials.gov Version History: A longitudinal mirror},
author = {Bret Gaulin},
year = {2026},
url = {https://huggingface.co/datasets/brbk/clinical_trials_history}
}
- Downloads last month
- 157