|
|
--- |
|
|
pretty_name: TheBlueScrubs-v1 (train) — fixed schema |
|
|
tags: |
|
|
- medical |
|
|
- healthcare |
|
|
- biology |
|
|
- text |
|
|
- pretraining |
|
|
- safety |
|
|
- classification |
|
|
- generation |
|
|
task_categories: |
|
|
- text-generation |
|
|
- text-classification |
|
|
language: |
|
|
- en |
|
|
license: apache-2.0 |
|
|
size_categories: |
|
|
- 10B<n<100B |
|
|
dataset_info: |
|
|
features: |
|
|
- name: text |
|
|
dtype: string |
|
|
--- |
|
|
|
|
|
# mkurman/TheBlueScrubs-v1-fixed |
|
|
|
|
|
## What is this? |
|
|
|
|
|
**TheBlueScrubs-v1-fixed** is a maintenance fork of the upstream [TheBlueScrubs/TheBlueScrubs-v1](https://huggingface.co/datasets/TheBlueScrubs/TheBlueScrubs-v1) *train split* that resolves a schema bug in the `meta` column. |
|
|
In the original train files, some rows serialized `meta` incorrectly (appearing as the literal string `"dict"`). This fork **re-exports the entire train split without `meta` column**, preserving text field and values. |
|
|
|
|
|
- **Document count:** 11,080,331 texts (train) |
|
|
- **Tokens (upstream estimate across all splits):** ~20B tokens |
|
|
- **Sources:** Curated from SlimPajama/RedPajama (Common Crawl, C4, GitHub, Books, arXiv, Wikipedia, StackExchange) |
|
|
- **Quality signals:** per-text medical probability (0.8–1.0) + three 1–5 LLM-based scores (relevance, precision/factual detail, safety/ethics); oncology label covering ~11B tokens across the full corpus. |
|
|
|
|
|
> Upstream details: The Blue Scrubs is a large, curated medical corpus designed for clinical LLMs, filtered via a logistic-regression screen and then Llama-3.1-70B evaluation; clinician and external checks reported high concordance. An oncology classifier adds cancer labels at scale. |
|
|
|
|
|
--- |
|
|
|
|
|
## Why this fork? |
|
|
|
|
|
- **Fix:** Removes the `meta` column, unblocking usage with `datasets` streaming and dataframe backends. |
|
|
- **Scope:** Content is otherwise **unchanged** relative to upstream train split (same rows, fields, and values). |
|
|
- **Goal:** Provide a drop-in train split that **loads cleanly** in `datasets` without ad-hoc parsing workarounds. |
|
|
|
|
|
--- |
|
|
|
|
|
## Data fields (train) |
|
|
|
|
|
| Field | Type | Description | |
|
|
|---|---|---| |
|
|
| `text` | string | Raw medical text extracted from SlimPajama/RedPajama sources. | |
|
|
|
|
|
--- |
|
|
|
|
|
## Splits |
|
|
|
|
|
This repository publishes the **train** split only (11,080,331 documents). For methods, scope, and aggregate corpus statistics (including validation/test in the upstream project), see the original dataset card and paper. |
|
|
|
|
|
--- |
|
|
|
|
|
## How to load |
|
|
|
|
|
```python |
|
|
from datasets import load_dataset |
|
|
|
|
|
# streaming |
|
|
ds = load_dataset("openmed-community/TheBlueScrubs-v1-fixed", split="train", streaming=True) |
|
|
row = next(iter(ds)) |
|
|
row["text"] |
|
|
|
|
|
# non-streaming (if you have local storage/network bandwidth) |
|
|
ds = load_dataset("openmed-community/TheBlueScrubs-v1-fixed", split="train") |
|
|
ds.features |
|
|
|
