Add 3 Sentence Transformers training skills#136
Conversation
| "name": name, | ||
| "description": description, | ||
| "path": str(skill_md.parent.relative_to(ROOT)), | ||
| "path": skill_md.parent.relative_to(ROOT).as_posix(), |
There was a problem hiding this comment.
This should be the only "outlier" change in this PR: this fixes Windows support for generate_agents.py. Not a big deal, but nice to have fixed.
|
Thanks @tomaarsen . One things that strikes me from a high level is that there are three skills with a lot of duplicated references. Would it be feasible to achieve this in one skill? Note, progressive disclosure means that an agent will only ever load the references it needs for the task it is doing. Also, multiple skills can hit adoption because we're asking users to do more. In concrete terms, we could move |
|
Working on a merged variant here: huggingface/sentence-transformers#3752 I'm also experimenting with making the SKILL.md tiny to force the agent to look at the references: https://github.com/huggingface/sentence-transformers/pull/3752/changes#diff-9edabb53f64ed74ad022f5ba6bcd5d52b0d5e56d9530f5243b6f3beb42ef1c8a
|
|
I've successfully merged the 3 into one, while making the agents load the correct files in my tests, for either of the 3 archetypes. Let me know your thoughts @burtenshaw !
|
burtenshaw
left a comment
burtenshaw
left a comment
There was a problem hiding this comment.
Nice. Tested in codex and claude code.
2 participants
Hello!
Pull Request overview
Details
This adds 3 training skills, which will also be hosted under https://github.com/huggingface/sentence-transformers and updated every time
sentence-transformershas an update via a syncing script. I'd like to get this PR in first, and then the syncing can take care of the updates.