Tooletta

Tool-call normalization for LLM dialects.

Tooletta is a small translation layer between model-specific tool-call formats and a canonical ToolCall schema. It is designed to fit the annoying part of model training, distillation, evals, and agent infrastructure: OpenAI, Qwen/Hermes, Mistral, Kimi, DeepSeek, and other model families do not agree on the text they emit for the same semantic tool call.

The center of gravity is Tinker-style training workflows. Tooletta is standalone, but its API is meant to stay easy to slot next to Tinker-compatible renderer, dataset-prep, and distillation code.

teacher trace / model output -> dialect parser -> ToolCall IR -> target renderer -> student format

Why

Tool calls are semantically simple and syntactically annoying. A model may emit <tool_call>{...}</tool_call>, another may use [TOOL_CALLS]..., another may expose OpenAI-compatible JSON objects, and training code still needs one reliable representation.

Tooletta's bet is simple: parse supported dialects at the boundary, keep a tiny canonical IR in the middle, and render only when you need model-specific text again.

That shape is directly inspired by Tinker's cookbook renderers: parse model-specific responses into canonical messages/tool calls, then render the canonical structure into the target model's format. Tooletta starts smaller on purpose: string-level tool-call parsing and rendering first, with tokenizer-aware training helpers considered only after the IR and dialect contract are solid.

Tinker Compatibility

Tinker compatibility is a first-class design target. Tooletta should make it boring to normalize tool calls before they enter a Tinker-style renderer or dataset builder, and to translate tool calls from one model-family dialect into the target format a student model expects.

The package does not depend on Tinker today and does not claim full Tinker renderer parity yet. The near-term goal is to keep the canonical ToolCall contract and parse/render APIs aligned with that style of workflow while the richer training helpers earn their tests.

The current boundary is intentionally boring: ToolCall.to_openai() returns the OpenAI/Tinker-style function-call object, and ToolCall.from_openai(...) accepts both plain mappings and object-style calls with .function.name / .function.arguments.

Scope

Tooletta currently handles:

• parsing supported tool-call dialects into ToolCall objects
• rendering ToolCall objects into supported target dialects
• best-effort dialect auto-detection
• custom dialect registration
• a stdin/stdout CLI for normalization and format conversion

Tooletta does not yet handle:

• tokenizer-specific token boundaries
• loss-mask generation
• full chat-template rendering
• streaming parser state
• full training or distillation pipeline orchestration

Install

After the first release:

uv add tooletta

For local development:

uv sync

Quick Start

from tooletta import parse_tool_calls, render_tool_calls

text = '<tool_call>{"name":"search","arguments":{"query":"tool calling"}}</tool_call>'

calls = parse_tool_calls(text, dialect="hermes")
print(calls[0].name)
print(calls[0].arguments)

print(render_tool_calls(calls, dialect="kimi"))

## Calling: search
{"query":"tool calling"}

CLI

Normalize a Qwen/Hermes-style tool call to canonical JSON:

printf '<tool_call>{"name":"search","arguments":{"query":"python"}}</tool_call>' \
  | uv run tooletta --from hermes

Render to another dialect:

printf '<tool_call>{"name":"search","arguments":{"query":"python"}}</tool_call>' \
  | uv run tooletta --from hermes --to kimi

Built-In Dialects

parse_tool_calls(..., dialect="auto") uses a stable built-in dialect order. Custom dialects are supported for explicit parsing/rendering by name; keep auto-detection for unambiguous built-ins.

Development

uv run ruff check .
uv run ruff format --check .
uv run mypy src tests
uv run pytest
uv run pre-commit run --all-files

Run the live Tinker smoke only when you intentionally want a real API call:

TOOLETTA_RUN_LIVE_TINKER=1 \
  uv run --env-file .env --with 'tinker>=0.9.0' pytest tests/test_live_tinker_smoke.py -q

By default it uses meta-llama/Llama-3.2-1B, rank 1, and a single training datum built from Tooletta-rendered tool-call text.

Status

Pre-alpha. The first goal is a fast, dependency-free runtime core that nails the IR, parser behavior, renderer behavior, and dialect plugin surface before growing a larger model-format zoo or training-specific APIs.