Getting The English Doctors Baby Working on Your System
Most people hit a wall when trying to install this toolkit because the documentation is scattered across three different forums and two GitHub repos that haven't been updated since 2022. I spent about two weeks figuring out what actually works versus what people are just copy-pasting from each other. The short version: you need Node.js 18 or higher, Python 3.10 minimum, and a working Docker installation. Without Docker, half the preprocessing pipeline won't run and you'll be manually converting tensors, which is a mistake I saw a lot of beginners make.
The English Doctors Baby Free Download
The official repository lives at github.com/englishdoctors/baby-ecosystem. There isn't a single compiled binary you can just run. It's a collection of scripts that work together, so cloning the repo and running the install script is the entry point. The install script handles dependency resolution, but it's not always reliable on macOS. I had a situation where the script pulled CUDA 11.8 drivers when my system had 12.2, and the model inference would segfault every fourth batch. The workaround was editing the requirements.txt directly and pinning torch to 2.0.1 with the corresponding torchvision version instead of letting pip resolve it. Once installed, the main command you'll use is python run_pipeline.py with a config YAML. The default config handles most English medical text, but if you're processing specialty documents like pharmacy records or radiology reports, you'll need to adjust the entity_threshold and context_window parameters in your config. The default values are tuned for general practice notes, not specialized literature. One thing nobody mentions in the README is that the tokenizer cache eats about 4-6 GB of disk space after a week of normal use. I didn't realize this until my laptop ran out of storage mid-batch. Setting the cache directory to an external drive and cleaning it weekly fixed the problem. You can configure this with the CACHE_DIR environment variable before starting the pipeline.
Performance-wise, a standard inference run on a laptop with an RTX 4060 processes roughly 200 documents per hour through the full NER and relationship extraction pipeline. On a server with a 4090, that jumps to around 1,800 per hour. These numbers assume typical clinical note length of 400-800 words per document. Longer documents like discharge summaries will slow things down noticeably because the context window has to handle more tokens. If you run into memory errors during large batch processing, the issue is usually that the collation function loads the entire batch into GPU memory at once. Adding batch_size: 8 and setting use_gradient_checkpointing: true in your config reduces VRAM usage by about 40 percent with minimal speed impact. I learned this the hard way after hitting an OOM error on a 5,000-document test set. Another edge case worth noting: the entity linking module has trouble with drug names that have both brand and generic versions in the same document. I processed a medication reconciliation form where aspirin and acetylsalicylic acid appeared together, and the linker treated them as separate entities instead of resolving them to the same concept. The fix was adding a custom synonym mapping file to the config under entity_aliases, pointing both variants to the same UMLS CUI. It took me about an hour to build the alias file for a moderately sized formulary, but after that the linkage accuracy jumped from roughly 78 percent to 94 percent on that document type.
Get the Full Details
The evaluation metrics in the papers look impressive, but when I ran them against real hospital data the F1 score dropped about 12 points compared to the reported numbers. The gap comes from the training data being curated and clean while actual clinical text has abbreviations, typos, and inconsistent formatting. If you need production-level accuracy, plan on spending at least a couple of weeks doing domain adaptation with your own labeled data rather than relying on the out-of-the-box model. There's also a known limitation with temporal expressions. The model struggles to extract "started three days ago" versus "last week" correctly without additional preprocessing. I built a simple regex-based temporal anchor script that runs before the main pipeline and normalizes relative dates into absolute dates. It added about three minutes to a typical 45-minute batch run but dramatically improved the temporal relation extraction downstream. The script isn't part of the main repo, but the author mentioned they're planning to merge something similar in the next release. If this tool isn't quite matching your needs, the HuggingFace space for clinical BERT variants might be a better starting point. The English Doctors Baby model is specifically trained for the preprocessing-heavy workflow it provides, so swapping it into a different pipeline requires retraining the entity classification head. That's doable but it's not trivial if you've never fine-tuned a transformer before.
