Overview
Key Highlights
Problem statement
Even frontier LLMs struggle to solve complex tasks that require tools. While the web offers an abundance of information, there are not that many datasets for training agents to solve problems with tools. Designing datasets at scale is not a trivial task
Approach
We have a unique team of diverse scientists and engineers who uses sophisticated GenAI processes to design datasets for training LLMs to solve difficult problems. Our datasets are 100% validated proven to lift performance
Results
We are providing high end datasets to the Frontier LLM companies. Our training data follow the terminal bench tbench.ai format and go through rigorous validation and testing. In a batch of 1000 tasks one problematic training point can ruin the results. We guarantee the quality of our data.
Project Overview
Details & Results
Domain
LLM Reinforcement Learning training
Tech Stack
- Tbench.ai, OpenAI, Anthropic, Google, LLama, Docker
Problem Statement
- Create tasks that a language model can perform with the use of tools (databases, planners, scientific software, etc)
Solution
- Generate task descriptions and docker files that contain the tools. Provide a grader and a solution to the problem. Validate the difficulty of the problem and prove that the LLM cannot cheat in order to hack the reward function.
Outcomes
- High quality datasets and trained models that demonstrate the lift in the performance