Sounds interesting! I'm trying to train a model but it's still "processing" after a bit but fine-tuning takes a while I get it. I'm having trouble understanding how it's inferring schema. I used a sample dataset and yet the sample inference curl uses a blank json?
How do I know what the inputs/outputs are for one of my models? I see I could have set the response variable manually before training but I was hoping the auto-infer would work.
Separately it'd be ideal if when I ask for models that you seem to not be able to train (I asked for an embedding model as a test) the platform would tell me it couldn't do that instead of making me choose a dataset that isn't anything to do with what I asked for.
All in all, super cool space, I can't wait to see more!
I'm a former YC founder turned investor living in Dogpatch. I'd love to chat more if you're down!
1. Depending on your dataset the training could take from 45 mins to a few hours. We do need add an ETA on the build in the UI.
2. The input schema is inferred towards the end of the model building process, not right at the start. This is because the final schema depends on the decisions made regarding input features, model architecture etc during the building process. You should see the sample curl update soon, with actual input fields.
3. Great point about upfront rejecting builds for types of models we don't yet support. We'll be sure to add this soon!
We're in London at the moment, but we'd love to connect with you and/or meet in person next time we're in SF - drop us a note on LinkedIn or something :)
Thanks for the great feedback! We've added a `baseline_deployed` status where the agents create an initial baseline and deploy it so you have something to play around with quickly. This is why you're seeing a blank json there. Once your final model is deployed, it creates an input and output schema from the features used for the model build :)
Product seems cool. But can you help me understand if what you are doing is different from the following:
> you put a prompt
> Plexe glorifies that prompt into a bigger prompt with more specific instructions (augmented by schema definitions, intent and whatnot)
> plug it into the provided model/LLM
> .predict() gives me the output (which was heavily guardrailed by the glorified prompt in the step 2)
Great question, and yes, it's quite different: Plexe generates code for a pipeline that processes your dataset (analysis, feature engineering, etc) and trains a custom ML model for your use case. When you call `.predict()`, it is that trained custom model that provides the response, not an LLM. The model is also hosted for you, and Plexe takes care of MLOps things like letting you retrain the model on new data, evaluating the model performance for you, etc. Using custom specialised models is generally more effective, faster and cheaper compared to running your predictions through an LLM when you have a lot of data specific to your business.
In the demo, you didn’t show the process of cleaning and labeling data, does your product do that somehow, or do you still expect the user to provide that after connecting the data source.
Great question, this is super important. The agents in the platform have the ability to do some degree of cleaning on your data when building a model (for example, imputing missing values). However, major improvements to data quality are generally not possible without an understanding of the data domain (i.e. business context), so you'll get better results if you "help" the platform by providing data in a reasonably clean state, answering the agent's follow-up questions in the chat, etc. By doing so you can give the agent better context and help it understand your data better, in which case it will also be more capable of dealing with things like missing values, misnamed columns etc.
This also highlights the important role of the user as a (potentially non-technical) domain expert. Hope that makes sense!
We have a data enricher feature (still in a beta mode) which uses LLMs to generate labels for your data. For cleaning and feature engineering, we use agents that automatically handle it for you once you've connected your data and defined your ML problem.
P.S. Thanks for the feedback on the video! We'll update it to show the cleaning and labelling process :)
very cool – I like how opinionated the product approach is vs. a bunch of disconnected tools for specialists to use (which seems more common for this space).
Thanks, we're pretty opinionated on "this should make sense to non-ML practitioners" being a defining aspect of the product vision. Behind the scenes, we've had quite a few conversations specifically about how to avoid features feeling "disconnected", which is always challenging at an early stage when you're getting pulled in several directions by users with different use cases. Happy to hear it came across that way to you.
Unfortunately we don't officially support image, video or audio yet - only tabular data for now. We do plan to add that capability at some point in the coming weeks depending on popular demand. Do you have any particular use case in mind?
Caveat: as a more technical user, you can currently "hack" around this limitation by storing your images as byte arrays in a parquet file, in which case the platform can ingest your data and train a CV model for you. We haven't tested the performance extensively though, so your mileage may vary here.
Sounds interesting! I'm trying to train a model but it's still "processing" after a bit but fine-tuning takes a while I get it. I'm having trouble understanding how it's inferring schema. I used a sample dataset and yet the sample inference curl uses a blank json?
curl -X POST "XXX/infer" \ -H "Content-Type: application/json" \ -H "x-api-key: YOUR_API_KEY" \ -d '{}'
How do I know what the inputs/outputs are for one of my models? I see I could have set the response variable manually before training but I was hoping the auto-infer would work.
Separately it'd be ideal if when I ask for models that you seem to not be able to train (I asked for an embedding model as a test) the platform would tell me it couldn't do that instead of making me choose a dataset that isn't anything to do with what I asked for.
All in all, super cool space, I can't wait to see more!
I'm a former YC founder turned investor living in Dogpatch. I'd love to chat more if you're down!
Thanks for the great feedback! To your points:
1. Depending on your dataset the training could take from 45 mins to a few hours. We do need add an ETA on the build in the UI.
2. The input schema is inferred towards the end of the model building process, not right at the start. This is because the final schema depends on the decisions made regarding input features, model architecture etc during the building process. You should see the sample curl update soon, with actual input fields.
3. Great point about upfront rejecting builds for types of models we don't yet support. We'll be sure to add this soon!
We're in London at the moment, but we'd love to connect with you and/or meet in person next time we're in SF - drop us a note on LinkedIn or something :)
Thanks for the great feedback! We've added a `baseline_deployed` status where the agents create an initial baseline and deploy it so you have something to play around with quickly. This is why you're seeing a blank json there. Once your final model is deployed, it creates an input and output schema from the features used for the model build :)
Product seems cool. But can you help me understand if what you are doing is different from the following: > you put a prompt > Plexe glorifies that prompt into a bigger prompt with more specific instructions (augmented by schema definitions, intent and whatnot) > plug it into the provided model/LLM > .predict() gives me the output (which was heavily guardrailed by the glorified prompt in the step 2)
Great question, and yes, it's quite different: Plexe generates code for a pipeline that processes your dataset (analysis, feature engineering, etc) and trains a custom ML model for your use case. When you call `.predict()`, it is that trained custom model that provides the response, not an LLM. The model is also hosted for you, and Plexe takes care of MLOps things like letting you retrain the model on new data, evaluating the model performance for you, etc. Using custom specialised models is generally more effective, faster and cheaper compared to running your predictions through an LLM when you have a lot of data specific to your business.
Really diggin this. Can't wait to try it out.
Thanks a lot! Excited for you to try it out and get your feedback :)
In the demo, you didn’t show the process of cleaning and labeling data, does your product do that somehow, or do you still expect the user to provide that after connecting the data source.
Great question, this is super important. The agents in the platform have the ability to do some degree of cleaning on your data when building a model (for example, imputing missing values). However, major improvements to data quality are generally not possible without an understanding of the data domain (i.e. business context), so you'll get better results if you "help" the platform by providing data in a reasonably clean state, answering the agent's follow-up questions in the chat, etc. By doing so you can give the agent better context and help it understand your data better, in which case it will also be more capable of dealing with things like missing values, misnamed columns etc.
This also highlights the important role of the user as a (potentially non-technical) domain expert. Hope that makes sense!
We have a data enricher feature (still in a beta mode) which uses LLMs to generate labels for your data. For cleaning and feature engineering, we use agents that automatically handle it for you once you've connected your data and defined your ML problem.
P.S. Thanks for the feedback on the video! We'll update it to show the cleaning and labelling process :)
very cool – I like how opinionated the product approach is vs. a bunch of disconnected tools for specialists to use (which seems more common for this space).
Thanks, we're pretty opinionated on "this should make sense to non-ML practitioners" being a defining aspect of the product vision. Behind the scenes, we've had quite a few conversations specifically about how to avoid features feeling "disconnected", which is always challenging at an early stage when you're getting pulled in several directions by users with different use cases. Happy to hear it came across that way to you.
How does it perform when build computer vision models?
Unfortunately we don't officially support image, video or audio yet - only tabular data for now. We do plan to add that capability at some point in the coming weeks depending on popular demand. Do you have any particular use case in mind?
Caveat: as a more technical user, you can currently "hack" around this limitation by storing your images as byte arrays in a parquet file, in which case the platform can ingest your data and train a CV model for you. We haven't tested the performance extensively though, so your mileage may vary here.
Great product!
Thank you! :)