benhamner

Can we add our own models or benchmarks?

Here's three main ways our users and customers use us:

1. Revenue Operations teams

Integrate Sumble's data programmatically to help with account scoring, territory planning, account qualification/disqualification, and CRM data cleanup. We provide feature matrices that feed ML models for large sales teams.

2. Individual AE's/SDR's

Many sales people have a small universe of named accounts that they go deep on. They use Sumble to understand the buying groups that exist within their target accounts, and which relevant technologies these groups use, and any relevant projects going on (e.g. data infrastructure migrations, cloud migrations, and GenAI projects can be critical signals for many of our customers)

For ongoing awareness of key changes within accounts, we work with our enterprise customers to define all the signals that are relevant to their sales plays, and send email/slack notifications when any of these signals happens in their accounts as well.

For sales reps with a larger universe of accounts (e.g. the SMB/commercial tier), they use us to filter out a lot of the noise in their territory and understand which accounts are real active businesses that are potential users of their product that they should spend time on.

3. Marketing

Marketers use us to figure out which accounts to focus on, and to spin up very targeted LinkedIn/Facebook/etc. campaigns to reach their most likely potential users and buyers

Thanks for the feedback!

The job functions we currently classify have been mostly focused by our early users/customers (companies building products/tools/infrastructure for data and software engineering teams), and handling the multilingual aspects of those across countries well.

We're aiming to extend this in two ways:

1. Adding job title and job description full-text search, to handle the long tail of usecases (in-flight project)

2. Extend the job function classification to the full universe of jobs that people can have

Thanks! Job descriptions are included in job post CSV exports, which is the current only path for that workflow.

We're planning to make that workflow much better in four ways this year:

1. Adding an API to make it easier to consume the data programmatically (next 2 months)

2. Enabling running LLM's on tabular results on Sumble directly that would enable pulling in job description context into the LLM call

3. Experimenting with an MCP endpoint, to see if that's helpful for these workflows as well

4. Experimenting with adding Sumble scoring models

Thanks! We're planning to add an API in the next two months, and exploring MCP alongside that

Thanks! We describe this as a knowledge graph because that's how we think about the structure in the data & is where we want to go.

Right now, we've focused on normalizing several key entities (e.g. organizations including parent/subsidiary relations, technologies, people, and job functions), and capturing the relations between these as well as additional useful metadata like location and industry.

From a backend implementation standpoint, this is currently implemented as structured relational tables for query performance and simplicity (e.g. count up all teams mentioning pytorch in job posts including rolling up across parent subsidiaries and sort by the biggest organizations descending).

Future direction here is TBD as we expand the sources that we cover and types of queries that can be computed across these sources.

There's been a lot of attempts at building high-quality public knowledge graphs that haven't hit escape velocity.

We're focusing on a structured, commercially relevant subset of the problem as a starting point to generate a critical mass of usage and funding that will enable us to build the bigger vision: a highly structured, up-to-date, and trusted repository of all the facts about the world that is easy to browse, query, and integrate programatically into all the relevant workflows (including for grounding LLM's)

Thanks! We'll take a look at that one

| Maybe specific but I want to filter by head count in job function (ie: find organizations that have 50-200 software engineers regardless of their total head count).

You're not alone! We've heard this from others as well, planning to add it soon

| 2. I don't know if this is happening because you are getting lots of traffic now, but each query takes 20-30 seconds which is unusable.

Thanks! What queries are you finding painful? Most should be under a second, there's some that are expensive though

Thanks! Our current coverage is focused on companies with a significant online presence (e.g. they've made job posts, people say they at the company, and/or they have a functional website).

Our goal is to have complete coverage for active companies and organizations in the world, and an understanding for companies that previously existed but are no longer active as well (these appear extensively in CRM's and add noise).

We prioritize expanding data coverage in areas that we hear are most useful from our current users and customers.

Thanks! Haven't prioritized something public facing on this front yet - what would you find most helpful?

Sumble | sumble.com | Data/Backend/Full-stack Engineering + first GTM | Remote-only North-America timezones

Sumble's applying language models (both small and large) to build high-quality datasets. Near-term our focus is company-related data that we integrate with customer's Salesforce data to make go-to-market operations more efficient. Long-term we want to become the first place you go to find high-quality data that lives outside your organization.

Currently a team of 7 engineers, including Kaggle/Google/Primer/Stack Overflow/Meta alumni.

Our tech stack includes Python, FastAPI, React/Typescript, GCP (postgres + alloydb + cloud run), regular expressions, and Pytorch/Gemma/Mistral.

Challenges we face include:

- Transforming noisy datasets + noisy models into high quality data products

- Making expensive analytics computations run efficiently

- Managing the complexity of an increasing number of data sources, models, and operations on top of these

If this excites you, apply here: https://wrkbl.ink/7tPh6M0

Oof, that can be scarily easy to do.

DBeaver by default has a warning you have to accept pop up if you try to run an UPDATE without a where clause. Good reason not to disable that warning.

This uses ~2-3k tweets per day for most days, which seems to be more than enough. According to https://twitter.com/WordleStats/status/1486021209015963649 there's about 250k daily tweets per Wordle right now, so this is about a 1% sample coming from whatever the Twitter search API returned when I ran that query.

The simulated distributions it's comparing to are based on 1000 runs per 5-letter word.

Anecdotally, 250 was enough to get it working for those simulated distributions, 100 and below it became increasingly noisier. A higher N would be nice, but I didn't spend more time optimizing the performance for the simulation code beyond what was needed to get this working.

Sorry about that! Just updated so today's guess is hidden by default (and you can click-to-unveil)

Sorry about that! Just updated so today's guess is hidden by default (and you can click-to-unveil)

Any way we can help at Kaggle? Is https://www.kaggle.com/datasets helpful for your work in connectomics?

We're working on that through Kaggle Datasets https://www.kaggle.com/datasets

We support data versioning, interactive web previews, seamless loading into hosted Jupyter notebooks (Kaggle Kernels), seeing/sharing analytic results built on the data version, and adding direct collaborators right now.

We don't support a data-oriented version of an "issue" or a "pull request" quite yet, but these needs are definitely on our radar.

I'd be interesting in understanding your usecase better. Can you send me an email? ben@kaggle.com

Strongly agree with this. We're aiming to make Kaggle Kernels/Datasets as reproducible as possible while being easy to use, with versioned data, versioned code, and versioned compute environments (the latter through Docker containers). https://www.kaggle.com/kernels

At the moment, we're focused on providing great support for the Python and R analytics/machine learning ecosystems. We'll likely expand this in the future, and in the meantime it's possible to hack through many other usecases we don't formally support well.

Code made public on Kaggle Kernels is currently required to be under an Apache 2.0 license.

Beyond that, for private work the non-legalese TL;DR is no, at least not beyond what's required for us to operate the service. I'll refer you to https://www.kaggle.com/terms, and copy one relevant section below. If your work is in the context of making submissions to a specific machine learning competition, then that competition may have bespoke exceptions to this as well (which would be detailed in the rules of the competition).

"For all User Submissions, you grant Kaggle a license to translate, modify (for technical purposes, for example making sure your content is viewable on an iPhone as well as a computer) and reproduce and otherwise act with respect to such User Submissions, in each case to enable us to operate the Services, as described in more detail below. You acknowledge and agree that Kaggle, in performing the required technical steps to provide the Services to our users (including you), may need to make changes to your User Submissions to conform and adapt those User Submissions to the technical requirements of communication networks, devices, services, or media, and the licenses you grant under these Terms include the rights to do so. You also agree that all of the licenses you grant under these Terms are royalty-free, perpetual, irrevocable, and worldwide. These are licenses only — your ownership in User Submissions is not affected."

You can also start a new Jupyter notebook session on any of these datasets with a click (click "New Kernel"), and then accelerate your analysis by attaching a GPU to the session with another click (for applications a GPU helps, e.g. training Tensorflow models on image data)

The competitions we host (https://www.kaggle.com/competitions) are supervised and always have a target we can create a numeric leaderboard on, but the public datasets (https://www.kaggle.com/datasets) are used for everything under the sun.

There's some supervised ML use of those, and a lot more open-ended exploration, visualization, cleaning, clustering, language modeling, etc.

All of the above, and more! One thing I'm really excited about that we're about to release is a much better explorer for tabular data (automated histograms, sorting/filtering/showing the data, and the like).

We also encourage sharing analytics code and visualizations that users create on the data back to the community. For example, see all these visualizations and insights in StackOverflow's developer survey data linked from https://www.kaggle.com/stackoverflow/stack-overflow-2018-dev...

Thanks for the feedback! Totally agree this could be clearer

Ben from Kaggle.

Open up the ~50 different individual datasets linked in separate tabs, and then quickly flip through all of them trying to get a sense of what each one is.

That experience will demonstrate one of the main challenges we're aiming to solve by making Kaggle Datasets your default place to publish data online (https://www.kaggle.com/datasets)

For those interested in a practical, hands-on course, we just released one at Kaggle https://www.kaggle.com/learn/overview

I shared this because another public data portal that I don't think has changed in years ended up at the top of HN. Kaggle Datasets has grown by over an order of magnitude in the past year, and jumps in scale fundamentally changes the utility of community products like this

We joined Google via acquisition one year ago, and Kaggle Datasets has grown from 450 datasets to over 13,000 in that timeframe. We are firmly committed to supporting and growing this platform.