README.md

---
license: cc-by-4.0
language:
- en
pretty_name: PubMed Embedding Vectors
tags:
- pubmed
- biomedical
- embeddings
- qdrant
- parquet
- retrieval
size_categories:
- 10M<n<100M
configs:
- config_name: neuml_pubmedbert_base_embeddings
  data_files:
  - split: train
    path: data/neuml_pubmedbert_base_embeddings/*.parquet
- config_name: qwen3_embedding_0_6b
  data_files:
  - split: train
    path: data/qwen3_embedding_0_6b/*.parquet
---

# PubMed Embedding Vectors

This dataset contains embedding vectors generated from local PubMed title and abstract text.
It is designed for biomedical retrieval and nearest-neighbor research.

The public files intentionally do not include PubMed titles, abstracts, or full text.
Rows contain PMIDs, embeddings, hashes, and lightweight metadata so researchers can join
against their own authorized PubMed mirror or the official NCBI/PubMed services.

## Configs

| Config | Model | Dim | Rows | Qdrant collection |
| --- | --- | ---: | ---: | --- |
| neuml_pubmedbert_base_embeddings | `NeuML/pubmedbert-base-embeddings` | 768 | 28,460,827 | `pubmed_emb_neuml_pubmedbert_base_embeddings_ddbc790c` |
| qwen3_embedding_0_6b | `Qwen/Qwen3-Embedding-0.6B` | 1024 | 28,460,827 | `pubmed_emb_qwen_qwen3_embedding_0_6b_cdca07b0` |

## Columns

- `pmid`: PubMed identifier.
- `embedding`: fixed-size embedding vector for the selected config.
- `text_sha256`: SHA-256 of the local title+abstract text used for embedding, when available.
- `pub_year`, `pub_month`: publication date metadata, when available.
- `raw_token_count`, `used_token_count`, `was_truncated`: embedding input token metadata, when available.

## Usage

```python
from datasets import load_dataset

repo_id = "aaekay/pubmed-embedding"
ds = load_dataset(repo_id, "qwen3_embedding_0_6b", split="train", streaming=True)
row = next(iter(ds))
print(row["pmid"], len(row["embedding"]))
```

Use the PMID to retrieve citation details from PubMed:

```python
pmid = row["pmid"]
url = f"https://pubmed.ncbi.nlm.nih.gov/{pmid}/"
```

## Source And Redistribution Notes

- Source records come from a local PubMed baseline/update mirror.
- NLM notes that PubMed abstracts may be protected by third-party copyright, so this
  dataset excludes article titles, abstracts, and full text.
- The generated embedding dataset is released as `cc-by-4.0`; upstream PubMed records
  and embedding models remain subject to their own terms.

Relevant upstream documentation:

- PubMed download page: https://pubmed.ncbi.nlm.nih.gov/download/
- NLM copyright information: https://www.nlm.nih.gov/databases/download.html
- Hugging Face large repository guidance: https://huggingface.co/docs/hub/storage-limits

## Manifest

Export metadata, shard checksums, and source collection details are stored in
`metadata/manifest.json`. The public schema is stored in `metadata/schema.json`.