Fine-Tuning as a Service (Vintage Data / SYNTH Methodology)

This entry captures the synthetic data methodology embedded in Fine-Tuning as a Service. See that page for the full platform evaluation.

SYNTH Methodology

  • Fully synthetic generalist training environment — trains SOTA small models without organic data
  • Used by Step-Fun (deep-research model) and multiple academic projects
  • Moving to agentic use cases requires qualitative step up: synthetic generation must actively create and solve real problems (not just verbalize existing data)

Why Fine-Tune Over Prompt

  1. Better performance on smaller models via specialization
  2. Accuracy on formal rules (math, structured output, rule-based systems)
  3. Style customization without ad-hoc prompts
  4. Built-in I/O structuration (markdown-delimited rubrics)
  5. Effective token use at scale

The Agentic Data Gap

Early agentic traces had ~30% faulty tool calls (bad JSON). Getting reliable agentic behavior requires training on high-quality agentic traces, not just prompting general models.