Introduction
Models can perform well on general benchmarks but fail in specific workflows. They may call tools incorrectly, miss required output formats, or fail on domain-specific reasoning.
SYNTH improves model performance by generating task-specific training data grounded in source files, such as PDFs and Parquet files. With focused data, a small model can reach a target benchmark result with fewer training tokens. SYNTH currently supports tool calling, multiple-choice questions, and question-answer pairs with support for mathematical operations, multi-turn reasoning, and harness environment datasets on the way.
-
Getting started
Install SYNTH and run your first generation job.
-
On-premise vs Cloud
Choose the deployment model that fits your data and operational requirements.