Baseten Launches Network of Vetted RL Environment Companies

Why Baseten is bringing in partners

Baseten says high-quality environments are critical to successful post-training. The company states it has built a science around creating RL environments and has applied it in each of its post-training engagements, naming Harvey, Lovable and Gamma among many others.

The work is demanding, Baseten says. It requires care, smart researchers, good taste and a lot of time spent looking at data. To scale it, the company says it needs more teams with that expertise, and the Network is meant to bring those teams to customers. Each environment is based on the customer's own data and domain.

Baseten says it spoke to a large number of RL environment companies, audited them and shortlisted heavily before choosing its partners. The company says it has deep trust in referring customers to them as an additional partner on their post-training journey.

How an engagement works

Baseten will refer customers to partners that can act as an extension of its applied research team. The process starts with scoping:

• Once Baseten identifies a project that could be a good fit, it goes through deeper scoping, discovery and calibration.
• Referrals are not automatic. Baseten says it does not want to refer a legal customer to a partner simply because that partner has done legal tasks before. The main question, it says, is whether it trusts the partner's process and can map it onto what the customer needs. Domain experience can matter in some cases, but is not the primary criterion.
• The partner then works directly with the customer, takes their real-world data and builds the specific environments they need. Depending on the engagement, this can also include creating tasks, reviewing tasks, creating rubrics and providing the resources to produce data within those environments.

Baseten still handles training and inference. Once the environments are ready, its training team trains the model and deploys it for the customer on Baseten's production inference platform.

Customer-specific environments and open-source plans

Baseten argues that building environments for specific customers, rather than for frontier labs, has its own advantages. A customer often arrives with data to seed environments, so it does not have to be generated synthetically. Customers also care about only one environment, which makes a scalable system for turning data and feedback into environments a worthwhile investment that they can use in perpetuity. Frontier labs, by contrast, typically want only a few thousand examples per task and so do not invest in such a system for each environment, according to the company.

Baseten also says it has spent significant time developing systems for auditing the quality of RL environments, including by using its training expertise. Soon, it plans to open-source a set of frontier-quality environments, many developed with partners in the network, together with details of how it audits them and how it trains on them. The company says it plans to use them for research, for standardizing evaluations, and as a real-world intelligence index representing information not available through current public benchmarks.

What it means for customers

For companies working on custom models, the Network adds an option for obtaining environments built on their own data and domain without building that capability in-house. Baseten says that, paired with its Training and Inference offerings, the network should help more customers reach value faster with their custom intelligence. Customers interested in connecting with Baseten's applied research team and the environment companies can reach out through the company's research support contact page.