[Feature Request] Support for tool / function calls

[leonardmq]

Is your feature request related to a problem? Please describe.

Fine-tuning for a Task that requires tool / function calls (as in these examples) is currently not supported.

Checks

• I've searched the docs for a solution
• I've searched for existing Github issues

Describe the solution you'd like

A rough idea of what that would likely involve:

• Task-level (or Project-level) defined tools where we define the tools and their schemas, etc.
• The Runs would display tool calls in some way and probably allow editing to correct incorrect tool calls

Describe alternatives you've considered

Not using tools 😄

Additional context

It seems like fine-tuning would be helpful for cases where tool usage depends on nuanced / semi-subjective context. For example, in a RAG setup, a model may need to decide whether to query a vector database or generate an answer directly based on context. Fine-tuning could help improve when and how it makes this decision.

Found this cookbook by OpenAI on fine-tuning for Function Calling: https://cookbook.openai.com/examples/fine_tuning_for_function_calling

[scosman]

Agreed. What's the best open standard for tools? MPC?

[quenbyako]

@scosman Currently, MCP is close to the unofficial standard of function calling (similar to openai api — not official, but a de-facto standard). So feels like it's way more worth to implement MCP here.

There is a different examples of how MCP could be supported, i found that it works well with llama's in oterm, maybe you can take a look there as an example of implementation.

But in any way, +1, could be extremely cool to implement it in Kiln, i found that this tool is amazing in terms of no-code fine tuning. So if you'll add MCP — it will be the best tool on the market 😍