stanley-nv/tool-routing-data

Tool Routing Data

Data for training and evaluating ToolRouter, a tiny open-set model that picks the right tool for a user query given tool names and descriptions. Two kinds of data:

Config	What	Splits (rows)
routing	user queries labelled with one of 7 priority tools (wiki, websearch, weather, news, files, shell, other)	train 3,927 · validation 805 · test 210 · test2 330
tools	tool catalogs: a tool name, two differently-worded descriptions and 9–14 example queries per tool; the tools in validation and test are never in train (open-set evaluation)	train 179 tools (2,409 queries) · validation 14 (140) · test 10 (100)
synthetic	8,665 template-generated routing queries (slot-filled templates with typo/case noise); low quality, used as light regularisation	train

Also priority_tools.json: names, training descriptions and one differently-worded evaluation description for the 7 priority tools.

from datasets import load_dataset
routing = load_dataset("stanley-nv/tool-routing-data", "routing")      # DatasetDict: train / validation / test / test2
tools   = load_dataset("stanley-nv/tool-routing-data", "tools")        # train / validation / test

Fields

routing/*.jsonl, synthetic/templates.jsonl

field	meaning
query	the user message
label	one of wiki, websearch, weather, news, files, shell, other
source	handwritten (written directly by the Claude Code assistant, not crowd-sourced), agent_gen_1..9 / agent_independent (written by Claude sub-agents), synthetic_template
style	writing style / theme the author was asked for: terse_keywords, long_polite, casual_typos, software_it_work, everyday_life, hard_boundary, files_vs_shell, other_boundary, web_news_wiki_weather, handwritten_seed, held_out_test, held_out_test2, template

tools/*.jsonl: name (snake_case), descriptions (list of 2 strings), queries (list of example user queries for that tool), source (handwritten or agent).

Labelling conventions (routing)

• wiki — stable encyclopedic knowledge: definitions, history, biographies, science/maths concepts, geography, how something works conceptually.
• websearch — live, commercial, local or practical web information: prices, reviews, comparisons, shopping, how-to/troubleshooting, software docs and latest versions, directions/hours, current office-holders and live status, scores/schedules, downloads, travel/visa procedures.
• weather — forecast, temperature, rain/snow, wind, humidity, UV, air quality, severe-weather alerts for a place/time.
• news — recent events, headlines, breaking news, updates on developing stories, announcements, daily briefings.
• files — operations on the user's local files/folders: open, read, show lines, list, find, search inside, create/write, edit, move/copy/delete, summarise a local document.
• shell — executing commands / system administration: run scripts, tests, builds, git, installs, docker, processes, services, disk/memory/GPU/network status, permissions, env vars.
• other — no tool needed: greetings, chit-chat, thanks, opinions, advice, creative writing, math, translation, writing/explaining code, rewriting pasted text, brainstorming, questions about the assistant.

Provenance — please read

• There is no real user traffic and no human-written query in this dataset. Everything was written by an LLM: the Claude Code assistant (Anthropic Claude) in an interactive session with the dataset's author (source: handwritten, i.e. "authored directly, not by sub-agents"), Claude sub-agents with prompts specifying a style and the conventions above (source: agent_*), and a template generator (synthetic_template).
• Labels were not human-verified. Boundary cases (wiki vs websearch vs news; files vs shell; other vs wiki) are judgement calls under the conventions above; expect label noise (we estimate several percent on the hard-boundary styles).
• routing/test (210) and tools/test (10 tools) were written by the assistant that also wrote training data (stylistic overlap is likely) and were looked at during development. routing/test2 (330) was written by a separate agent for the purpose of an independent test, then filtered: every query that matched any training/validation/test query after lower-casing and stripping punctuation was removed (90 of 420), so it is wiki-light (23 wiki queries). tools/validation was used for model selection; tools/test for final reporting.
• Leakage control: across all routing splits and synthetic, queries are de-duplicated after normalisation (lower-case, non-alphanumerics removed) with priority test > test2 > validation > train. Tool names are unique across the tools splits. Related-but-different tools can still be semantically close (e.g. dictation vs text_to_speech).
• Sanitised: file paths use /home/user/..., email addresses use example.com, people are fictional or well-known public figures, and companies/products appear only as public topics. A scan for emails, phone numbers, tokens, usernames, and employer- or organisation-specific names found none.

Intended use and limitations

• Training/evaluating small tool routers and intent classifiers; benchmarking open-set generalisation to unseen tool descriptions.
• English only; short single-turn queries; one tool per query; the tool catalogs cover ~200 invented tool domains (productivity, dev/IT, travel, finance/health/education, media/science, devices/government/HR) and are not real API schemas.
• Terms: because the text was generated with Claude, review Anthropic's terms and usage policy for your use (in particular regarding training other AI models) before commercial use. The data are released under CC-BY-4.0 by the author; change/confirm the licence before wider publication.

Validation

python validate.py checks the schema, labels, duplicates inside and across splits, and prints the split statistics.

Citation

@misc{toolroutingdata2026,
  title  = {Tool Routing Data},
  author = {Stanley},
  year   = {2026},
  url    = {https://huggingface.co/datasets/stanley-nv/tool-routing-data}
}