new in ml-intern: you can now actually see what's going on inside added native metric logging + trackio integration. every training run the agent kicks off now has live curves you can watch in real time before it was kind of a black box. agent launches a job, you... wait, you hope. now you see exactly where it's learning, where it's not, and when something goes wrong transparency for free, basically. should i add something more?show more

@huggingface This is so surprising

@aksel timestamp on each action, an exportable log of actions, sometime it hang up and you can’t continue it properly.

black box training loops are one of the most frustrating parts of the research cycle. native metric logging that the agent itself can watch changes the feedback loop entirely. nice

Ser can you fix ml-intern. i bought the pro plan and i cant actually acccess any of the features of it.

lovely! Do you plan to allow using any subscriptions to use the mlintern (like opencode sub)?

we could... which ones do you want?

Thanks! I'd love to be able to use my Codex & Opencode subscriptions

@_akhaliq This is one of the most exciting projects in a long time. You’re already moving fast but I c a whole bunch of post training areas that this can be extended into I’m actively thinking about how to apply it to evolving my domain agents Thank you for building AND open sourcing it

The teams who got early access to this kind of visibility cut their debugging time in half. Real-time curves change how you catch drift before it becomes a wasted run.

Is it possible to change permissions? I was able to set them the first time but I can't find how to change the organizations visible to the token.

When I used it, it started a full training run and during that time the prompt was busy. I saw that loss was too high but couldn't steer it. When I stopped the prompt it didn't stop the training

This is a strong direction. The interesting part is not just automating research loops, it is making the outputs reviewable so teams can trace which paper, experiment, or citation actually changed the model behavior.

ml-intern's logging is cool but i'd rather just grep the logs myself than wait for another dashboard to render

I don't think you should add anything else yet i can using this ml-intern?

@_akhaliq What interface are you running this on? Is there a web version available? The cli is great but lacks observability

live curves are a huge relief for long runs. nothing kills productivity like finding out your agent spent three hours chasing a local minimum because you couldnt see the loss exploding in real time.

this looks amazing!!! I have to give it a try. Ive been making my own app to connect huggingface, arxiv, and google colab together as a single app to playtest training models on. Is ml-intern like that? I keep seeing it but I don't know if what I'm making is stepping on toes a little.

I’m about to have so much fun at my job

Per-step vs per-trial granularity is where dashboards usually break. If trackio defaults to per-step, a sweep-level summary view saves the dashboard once you fan out runs.