|
--- |
|
task_categories: |
|
- text-generation |
|
language: |
|
- en |
|
size_categories: |
|
- 1M<n<10M |
|
--- |
|
## LitScan EPMC Subset |
|
|
|
This dataset is a subset of [afg1/epmc-oa-subset](https://huggingface.co/datasets/afg1/epmc-oa-subset), |
|
which itself comes from the [Europe PMC open access subset](https://europepmc.org/downloads/openaccess) of about 5.9 million articles. |
|
|
|
Here, we take the ~960 parquet files from the full OA subset and join them against a list of PMCIDs for articles found by [LitScan](https://rnacentral.org/help/litscan), |
|
which should discuss ncRNA for the ~9.6 million IDs searched from RNAcentral. The result is a collection of just over 1 million open access |
|
fulltext articles ostensibly about ncRNA. |
|
|
|
The primary use case for this is pre-finetuning on domain specific text. This idea of domain adaptation is similar to what |
|
NVIDIA have done with their [ChipNeMo model](https://research.nvidia.com/publication/2023-10_chipnemo-domain-adapted-llms-chip-design). |
|
|
|
We are planning to finetune some models on this dataset, probably TinyLlama, since it is quite quick to train. |
|
These will be useful for e.g. generating embeddings for RAG, or further downstream finetuning on tasks like summarisation. |
|
|
|
## Limitations |
|
|
|
The epmc-oa-subset parquet files are parsed from JATS, which does not always go entirely to plan. As a result, there are likely to be some |
|
articles with missing text, or strange tags left in. These should be quite rare, but I can't guarantee they're not in there. |
|
|
|
LitScan itself also has some limitations, namely that there is quite a high false positive rate for those RNA IDs that are a bit generic. This |
|
means that while most of the articles in this dataset should be focused on RNA, there will be a significant minority that are about all sorts of |
|
other things, including but not limited to: concrete, female mice, recurrent neural networks. This is a very tricky problem to solve! |
|
|
|
|
|
|
