Feature Request / Question: Pretrained Checkpoints and Custom Time-Series Extension

[eripe970]

Question

Hi,

Thanks for sharing this project — really impressive work.
I was wondering if there are any pretrained models or checkpoints available for any of the stages (TSQA, M4 captioning, HAR-CoT, Sleep-CoT, ECG-QA-CoT)?

It would be great to start from an existing checkpoint instead of training everything from scratch.

Thanks!

Solution

N/A

Additional context

No response

Code of Conduct

• I agree to follow this project's Code of Conduct and Contributing Guidelines

[RealLast]

Hi @eripe970
Thank you so much!
Checkpoints for all models except TSQA are available here: https://huggingface.co/OpenTSLM

Thanks and best

[CornelioP]

Hello, guys!
First of all, thank you for your work. I tried to test your models and download the weights from Hugging Face but I'm a bit confused by the nomenclature you used. I don't understand which models are SP and which are Flamingo. For example in this case here, the folder name is Sleep Flamingo, but the model file name is Soft Prompt. It would be greatly appreciated if you could clarify which file I should download to test the flamingo model.

Best!

[eripe970]

Isn't it just like if the folder name ends with flamingo, then it's for flamingo, otherwise not. Soft prompt is the trained prompt embeddings that adapt the base model.

[eripe970]

@RealLast, any reason to why the TSQA checkspoints aren't uploaded?

[robsyc]

Hello, I've experimented with several of the pre-trained models and found them to be very over-fit to the 'autoregressive language modeling' training objective. ECG-trained models prompted with other medical data streams or only slightly different prompts than the ones used during training will consistently output gibberish even when in-context examples are provided.

Is there any possibility to additionally upload the TSQA- and M4-trained models? I'm hoping these models will have retained more of their generic in context learning abilities that make Flamingo so desirable.

Anyways, great research and interesting approach. I see this exploding very soon :)

[YvesSchoenberg]

Awesome research! +1 for uploading the TSQA and M4-trained models :)

[RealLast]

Hi everyone!
Thank you so much for your interest and for the kind words! We really appreciate it. Sorry for the late response, it has been a really busy couple of weeks recently.

We have now uploaded all model checkpoints to huggingface, including TSQA and M4:
OpenTSLM huggingface

Additionally, we have a branch with demo scripts showing how to use these checkpoints and the corresponding datasets. If you use these scripts and our checkpoints you should be able to reproduce our results:
Huggingface demo scripts

Please let us know if it works for you or if you run into any issues. Additionally, if you are building cool projects with OpenTSLM and/or want to collaborate, please reach out! 🔥

We are already working on the next-generation architecture of TSLMs and will release more powerful models with better generalization across datasets in the near future.

Best