- Specialization
- Language Models
- Business
Model Evaluationand 2 more business valuesLess
- Data Provision
- Model Development
- Market
- Software Development
- Ecosystem
- Not classified
- Customers
- Enterprises
More details
- Organization type
- Commercial Company
- Ecosystem
- Not classified
- Business
- Model Evaluation, Data Provision, Model Development
- Specialization
- Language Models
- Market
- Software Development
- Customers
- Enterprises
- Delivery models
- Professional Services, On Premise Software
- Revenue models
- Not classified
- Listing status
- Not classified
- Latest equity round
- Not classified
- Reported funding amount
- Not classified
About
The LLM Data Company, now presented as Paper Instruments, post-trains LLMs and builds evaluation, dataset, and reinforcement-learning infrastructure for frontier AI teams and applied-AI companies. Its differentiation is an infrastructure-first workflow for high-signal evaluations, aligned graders, fine-grained rubrics, versioned evals-as-code, and model-reward definition.
Market
Paper Instruments, formerly The LLM Data Company, competes in LLM evaluation and post-training infrastructure for frontier models, with an adjacent position in agent software and RL environments for knowledge work. Its differentiation is an integrated eval-to-training workflow: doteval combines eval authoring, aligned grading, versioning, run comparison, reward definition, and exportable RL training specifications, rather than focusing only on monitoring or standalone evaluation.
Their ideal customers are frontier AI labs and applied-AI companies, especially teams developing or post-training LLMs and complex agent or knowledge-work systems. The likely buyers and users are ML research, post-training/RL, evaluation, and agent-engineering teams that need bespoke tasks, rewards, benchmarks, and training environments.
At a Glance
Problem
The LLM Data Company addresses a central bottleneck in post-training: teams building models and agents need reliable, high-signal evaluations and reward definitions for complex, unstructured work. Without them, it is difficult to know whether a new checkpoint, model, or prompt actually improves performance, and reinforcement-learning methods such as GRPO or RFT cannot be trained effectively. The economic pain is primarily wasted research and engineering time, expensive iteration, and the risk of deploying a model that appears capable but performs poorly in real-world settings. The clearest high-stakes use case is specialized reasoning, particularly medicine, where general-purpose models can be sycophantic, insufficiently grounded, or too verbose for safe clinical assistance.
Product / Service
The company’s core offering is doteval, an eval workspace for creating high-signal evaluations, defining reward functions, and preparing reinforcement-learning environments. It is described as a Cursor-like interface for editing evals as code in a YAML schema, versioning them across checkpoints, generating AI-assisted diffs, and comparing runs. In practice, this lets AI teams turn bespoke tasks, rubrics, and graders into repeatable measurement and training loops rather than relying on manual testing.
The company also combines this tooling with research and model development. Its current site presents Paper Instruments as an open-source frontier-models company for knowledge work, while its research has produced Kos-1 Lite, a medical reasoning model designed to prioritize clinical accuracy, reduced sycophancy, triage, and deferral. The benefit is faster, more defensible model iteration and lower-cost specialization: Kos-1 Lite’s announcement claims leading HealthBench performance at a fraction of the serving cost of very large frontier models.
Market
The company competes in AI infrastructure, specifically LLM evaluation, post-training data, reward-function, and agent-development tooling. Its positioning sits between developer platforms for evaluating and monitoring models and specialized research infrastructure for reinforcement learning. The available competitor mapping names Databricks, Arize AI, and Braintrust as possible competitors, although The LLM Data Company differentiates itself by emphasizing evals-as-code, reward definition, bespoke environments, and direct work with frontier AI teams. Its current site also indicates expansion toward agent-first knowledge-work software and open-source frontier models.
The company appears to be early-stage but active rather than an established scaled vendor. Y Combinator lists it as a Spring 2025 company and says it works with frontier AI teams; third-party funding data reports a $500,000 seed round and a three-person founding team. It launched its Workspace for Evals in 2025 and announced Kos-1 Lite in March 2026, reporting a 46.6% score on HealthBench Hard. The evidence reviewed does not disclose revenue or named customer logos, so it is best characterized as an early commercialization or research-led company, not demonstrably revenue-scaled.
Founders & Leadership
Funding History
Y Combinator
Recent News
Startup Intros updated its profile of The LLM Data Company, describing it as a research lab focused on post-training data. The profile says the company works with frontier AI teams to create bespoke tasks, rewards, and environments for models.
The LLM Data Company announced Kos-1 Lite, its first medical reasoning model. The company reports that it achieved 46.6% on HealthBench Hard, described as state-of-the-art, while prioritizing clinical accuracy, triage, deferral, and compassionate bedside manner.
Active Roles
0No active roles right now.
Get notified when they postBusiness Model
The public materials indicate a B2B model serving frontier AI labs and applied-AI companies through bespoke evaluation datasets, tasks, rewards, reinforcement-learning environments, and related software such as doteval. Public pricing is not disclosed; the company directs interested teams to contact the founders for early access or evaluation-dataset help.