Baseten 推出 Baseten Environment Network

Today, we're introducing the Baseten Environment Network, a group of highly vetted RL environment companies that will build custom RL environments for our customers. 

High-quality environments are critical to successful post-training, and we’ve spent a lot of time building a science around RL environment creation. We've done this for each of our post-training engagements (Harvey, Lovable, Gamma, and many more). But it is manually intensive work: it requires care, smart researchers, good taste, and a lot of time spent looking at data. To scale that, we need more teams with that expertise. 

The Baseten Environment Network brings those teams to our customers to deliver high-quality RL environments based on each customer's unique data and domain. We believe that paired with Baseten’s Training and Inference, this network will help more customers reach value faster with their custom intelligence.

The Baseten Environment Network

We spoke to a large number of RL environment companies, audited them, and shortlisted heavily. We're proud to partner with some of the best RL environment companies on the market, and have deep trust in referring our customers to them as an additional partner on their post-training journey.

Loading embedded content…

How an engagement with the Network works

Baseten will refer customers to trusted partners who can act as an extension of our applied research team:

1. We’ll scope each engagement around the customer’s needs: Once we identify a project that could be a good fit, we’ll go through deeper scoping, discovery, and calibration. We don’t want to automatically refer a legal customer to a partner just because they’ve done legal tasks before. The main question is: do we trust their process, and can we map that process onto what this customer needs? Domain experience can matter in some cases, but it isn’t the primary referral criterion.

2. The partner will work directly with the customer. They’ll interact with the customer, take their real-world data, and build the specific environments the customer needs. Depending on the engagement, that can also include creating tasks, reviewing tasks, creating rubrics, and providing the resources to produce data within those environments.

3. Baseten still does the training and inference. Once the environments are ready, our training team will go about training the model and deploying it for the customer onto Baseten’s world-class production inference.

Own your RL environments for good

Making environments for specific customers, rather than frontier labs, also has its own benefits. For example, a customer often comes with their own data to seed environments, rather than having to synthetically generate it. Additionally, customers care about only one environment, so building a scalable system to turn data and feedback into environments is a worthwhile investment per customer. They can use this system in perpetuity, and it is worth the upfront cost, as opposed to frontier labs, which typically want only a few thousand examples per task and thus do not invest in a scalable system per environment.

Open environments coming soon

We have spent significant time developing systems for auditing the quality of RL environments, including leveraging our training expertise in order to assess their quality. Soon, we will open-source a set of frontier-quality environments, many developed in collaboration with partners in the network, along with how we audit them and how we train on them. We plan to use these environments for many things, such as research, standardizing evaluations, and as a real-world intelligence index that represents information not available through current public benchmarks.

If you're interested in connecting with our applied research team and these environment companies, reach out here: https://www.baseten.co/talk-to-us/research-support/