Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.veri.studio/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

Function calling is a strong fit for Veri’s current fine-tuning approach. If you want a model to emit predictable structured outputs, choose tools more reliably, or follow a strict schema, you can shape that behavior with datasets plus reward functions. Veri does this today through reward-based fine-tuning with GRPO.

​
Good Fits

Use this workflow when you want the model to:
  • return valid JSON
  • emit function names and arguments in a strict format
  • prefer tool usage over free-form answers
  • follow schema and formatting rules consistently

​
1. Build a structured dataset

Your dataset should include prompts plus the fields needed to evaluate correctness, such as: 
{"prompt": "Book me a flight to New York tomorrow", "expected_tool": "book_flight"}
{"prompt": "What's the weather in SF?", "expected_tool": "get_weather"}

​
2. Reward the behavior you want

Common reward signals for function-calling use cases include:
  • correct function name selected
  • valid argument schema
  • valid JSON or XML formatting
  • no extra text outside the structured response

​
3. Keep the task narrow at first

Start with one or two tools and a small schema. Once the model reliably follows the format, expand the tool surface.

​
Example Reward Ideas

  • +1.0 for the correct tool
  • +0.5 for valid structured syntax
  • -1.0 for malformed output
  • -0.5 for answering directly when a tool should be called
That kind of shaping usually works better than a single binary reward.

​
What Veri Supports Today

  • single-turn fine-tuning workflows
  • custom reward functions
  • flexible dataset inputs
  • downloadable Hugging Face checkpoints

​
Current Limits

The current specs call out a few important constraints:
  • multi-turn agentic RL is not the focus of the PoC
  • hosted serving is not part of the current product
  • function-calling quality still depends heavily on dataset quality and reward design
So the best fit today is: fine-tune for structured one-turn tool selection and schema-following behavior, then deploy the resulting checkpoint in your own serving stack

​
Where to Go Next