Clarify embedding generation in README and codebase

[julesjacobsen]

The curate-index command is not part of ELDER - clarify where this / these external dependencies are hosted:

https://github.com/monarch-initiative/ELDER?tab=readme-ov-file#creating-your-own-embeddings

In the paper you state that the ontology term relationships are used and produce the best results, yet its not clear how you generated these from the codebase. It could be using the curate-index commands (looks like it), but then there is this file https://github.com/monarch-initiative/ELDER/blob/7bb21744f1eaec37f82a314411b24711a514c4a0/src/pheval_elder/prepare/core/llm_descriptions/hpo_def_generator.py which appears to use an LLM to generate descriptions for an ontology term based on the id and label.

Is this still relevant to the project as described in the paper? If not, it should be deleted / moved into a different repo.

Lastly what were the exact commands and data versions used for each dataset you generated on Huggingface? This should be stated ideally on Hugging face as part of the data card e.g.

iQuxLE/ada-002_lrd_hpo_embedding
---
HPO version: v2025-08-11
Software: curate-index
Software version: ????
Software source: ???
Command: curate-index index-ontology --index-fields "label,definition,relationships,aliases" --model ada002

[iQuxLE]

Hi Jules,

thank you very much for your comments!

curate-index: https://github.com/monarch-initiative/ELDER/blob/master/src/pheval_elder/curate_index_cli.py

In the paper I used the commands inside curate-gpt. Nothing from this repo, everything in this repo I was trying to built to make it easier to use without cloning and getting into 'curate-gpt'. There might be issues which I am very happy to resolve. But best is always to use 'curate-gpt'.

This is the command I used to create the embeddings!

The

Iinitiative/ELDER/blob/7bb21744f1eaec37f82a314411b24711a514c4a0/src/pheval_elder/prepare/core/llm_descriptions/hpo_def_generator.py which appears to use an LLM to generate descriptions for an ontology term based on the id and label.

can be taken out.

This was for an extra experiemtn which in the end I did in a very different way and it also ended up to not make things better, however still it was very high scoring and beat ontology-based methods. Let me try and find all the things and put them together this might take a day or two.

Thanks for making me aware of putting version numbers, this I will do as well.