README.md

---
pretty_name: PubMed Research Classifier Labels
license: other
license_name: embo-internal
license_link: LICENSE.md
task_categories:
  - text-classification
language:
  - en
tags:
  - pubmed
  - openalex
  - research-classification
  - bibliometrics
  - embo
size_categories:
  - 10M<n<100M
configs:
  - config_name: v1.0.0
    data_files: data/v1.0.0/labels.csv
    default: true
dataset_info:
  config_name: v1.0.0
  features:
    - name: PMID
      dtype: string
    - name: class
      dtype: string
    - name: probability
      dtype: float64
  splits:
    - name: train
      num_examples: 30426295
---

# PubMed Research Classifier Labels

Private lookup table of **research / non-research** labels for PubMed IDs.

| | |
|--|--|
| **PyPI package** | [`pubmed-research-classifier`](https://pypi.org/project/pubmed-research-classifier/) (**≥ 0.3.0** for Hub lookup) |
| **This dataset** | Hub release **v1.0.0** (labels scored with package **0.2.0** weights) |
| **Docs** | [Package README on PyPI](https://pypi.org/project/pubmed-research-classifier/) |

**Coverage (v1.0.0):** labels for **all OpenAlex works that have a PubMed ID (PMID)**,
drawn from the OpenAlex corpus snapshot used in this project, with PubMed-linked
works **up to April 2026**. Every distinct PMID in that slice is included
(**30,426,295** unique IDs). Works without a PMID are out of scope for this table.

This Hub dataset is versioned independently (`v1.0.0`, `v1.1.0`, …) so future
classifier / corpus refreshes can ship updated label tables without breaking
consumers pinned to an older revision.

| Field | Version |
|-------|---------|
| **Dataset release** | **v1.0.0** (this revision) |
| **Corpus coverage** | All OpenAlex works with PMID, up to **April 2026** |
| Label model (weights) | `pubmed-research-classifier==0.2.0` |
| Recommended client | `pubmed-research-classifier>=0.3.0` (`[cache]` extra) |
| Decision rule | `research` if P(non-research) &lt; **0.75**, else `non-research` (τ = 0.75) |
| Unique PMIDs | **30,426,295** |

---

## Dataset structure

| Column | Type | Description |
|--------|------|-------------|
| `PMID` | string | Numeric PubMed ID (no `pmid:` / URL prefix) |
| `class` | string | `research` or `non-research` |
| `probability` | float | Model estimate of **P(non-research)** |

Class counts in **v1.0.0**:

| `class` | Count |
|---------|------:|
| `research` | 24,113,963 |
| `non-research` | 6,312,332 |

Files:

```
data/v1.0.0/labels.csv
```

Later releases add `data/v1.1.0/`, `data/v2.0.0/`, etc., each exposed as a
dataset config with the same three columns.

---

## How to use with `pubmed-research-classifier` (≥ 0.3.0) — recommended

Install the optional cache extra and set a Hub token with **read** access to
this private dataset:

```bash
pip install "pubmed-research-classifier[cache]>=0.3.0"
export HF_TOKEN=hf_xxxxxxxx   # or: huggingface-cli login
```

```python
from pubmed_research_classifier import (
    load_label_cache,
    lookup_pmid,
    classify_pmid,
)

# First call downloads v1.0.0 (~1.2 GB) and builds a local DuckDB index
# under ~/.cache/pubmed_research_classifier/ (override with PUBMED_RC_CACHE_DIR).
cache = load_label_cache(revision="v1.0.0")
print(len(cache))  # 30426295

lookup_pmid("10006576")
# {"PMID": "10006576", "class": "research",
#  "probability": 0.009..., "source": "cache"}

# Cache hit → Hub row; miss → run the bundled MLP (needs classify() fields)
classify_pmid("10006576")
classify_pmid(
    "99999999",
    record={
        "title": "...",
        "abstract": "...",
        "pub_types": ["Journal Article"],
        "n_authors": 3,
        "n_refs": 12,
    },
)
```

Batch lookup (order-preserving):

```python
rows = cache.lookup_many(["10006576", "10047518", "99999999"])
# [dict, dict, None]
```

`classify(...)` without the cache still works for raw text / embeddings
(see the [PyPI README](https://pypi.org/project/pubmed-research-classifier/)).

### Monthly Hub refresh (maintainers)

Needs a token with **write** access. Merge a CSV of new PMIDs
(`PMID,class,probability`) into the previous table and upload a new revision:

```bash
pubmed-rc-publish-labels \
  --new-csv new_pmids.csv \
  --base-revision v1.0.0 \
  --new-revision v1.1.0 \
  --upload
```

Or from Python: `publish_label_revision(..., upload=True)`.

---

## How to load with 🤗 Datasets (without the package)

Requires a Hugging Face token with access to this **private** dataset.

```python
from datasets import load_dataset

# Default config = latest published table (currently v1.0.0)
ds = load_dataset(
    "EMBO/pubmed-research-classifier",
    token=True,          # or pass a token string / use HF_TOKEN
)

# Pin a release
ds = load_dataset(
    "EMBO/pubmed-research-classifier",
    name="v1.0.0",
    token=True,
)

print(ds["train"][0])
# {'PMID': '10006576', 'class': 'research', 'probability': 0.009...}
```

For large joins, prefer DuckDB / Polars on the CSV (or the package’s local
DuckDB file) rather than building a giant Python dict.

---

## Provenance

**v1.0.0** covers **all OpenAlex works sourced from PubMed (i.e. with a PMID) up to
April 2026** — one row per distinct PMID in that OpenAlex/PubMed-linked slice.

Labels were computed from OpenAlex-derived work tables (people works, life-science
author works, and a large ModernBERT-embedded OpenAlex shard) by:

1. Building per-PMID features (ModernBERT title/abstract embeddings + PubMed metadata).
2. Running `pubmed-research-classifier==0.2.0` in **embedding mode** on GPU.
3. Exporting the unique-PMID result table uploaded here.

Local mirror of this file (EMBO DGX):

`/raid/shared/apply-classify-article/data/pubmed-research-classifier_cached-results_0.2.0.csv`

---

## Versioning policy

| Hub dataset tag / config | Meaning |
|--------------------------|---------|
| `v1.0.0` | First EMBO release of the PMID label table (scored with classifier **0.2.0**) |
| `v1.x.0` | Additive / corrective corpus updates (same schema); use package **≥ 0.3.0** to load |
| `v2.0.0` | Breaking schema or new primary classifier generation |

Always pin `revision=` / `name=` (Hub) or `load_label_cache(revision=...)` (package)
in production code.

---

## License / access

**Private** EMBO dataset. Do not redistribute outside authorised Hugging Face
accounts/organisations. Contact EMBO data owners for access requests.