Borcherding/calculator-routing

Calculator routing

Synthetic typed-decision rows for one call an agent makes before each step: does this step need a calculation?

A calculation means producing a number by computing it: arithmetic, percentages, unit or currency conversion, date math, statistics, or code that crunches numbers. Looking a number up, quoting one a tool already returned, or numbers used as names and IDs do not count. Code and a calculator are treated the same: the question is whether something has to be computed, not which tool does it.

Rows use the LocalLLaMA/typed-decisions shape (state, questions, gold), so Unsloth Studio can fine-tune a Laya decision model on them directly.

Questions

question	type	answers
needs_calculation	yes/no	does the agent's next step need a calculation?
difficulty	score 0 to 2	none; trivial, done in your head; real working (several steps, big or awkward numbers, a formula)

gold holds a label and a probability for every answer, so the model learns soft targets. Clear cases sit near 0.03 or 0.97; rough estimates (0.8) and writing code about money without running it (0.2) sit closer to the middle.

State

Each state is an agent step, not just a chat message:

{
  "task": "You are a travel planning agent.",
  "tools": ["calculator", "web_search", "code"],
  "thread": [
    {"role": "tool", "name": "web_search", "text": "option 1: €142.1 per night in Porto; option 2: €98.4 per night in Porto"},
    {"role": "user", "text": "Book the cheapest for 4 nights and tell me the total."}
  ]
}

• task: one of 10 agent roles (shopping, travel, finance, coding, GPU cloud support, data analysis, ...).
• tools: one of 7 tool sets, including none. The label is about whether a calculation is needed, not whether a tool for it is available.
• thread: 0 to 2 earlier turns, sometimes a tool result, then the message for this step. In handoff_calc rows a tool returned numbers that now have to be combined; in handoff_nocalc rows the number is already there and only has to be reported or acted on.

Families

family	needs calculation	train	test	example
handoff_calc	yes	300	60	(after a search returned prices) What's the average of those?
conversion	yes	279	47	four miles in km?
arithmetic	yes	270	47	20 to the power of 3?
stats	yes	264	48	Normalise 80.05, 916, 910, 830, 245, 141, 98.7 so they add up to 1.
practical	yes	262	61	I need 548 g of protein a day and each shake has 10 g. How many shakes?
percent	yes	260	43	Which is the bigger discount: 12.5% off CHF 23009 or CHF 24,047 off?
trivial	yes	257	58	what's 13 + five
totals	yes	255	54	Shipping is CHF 96.90 per box and each box holds 3 coffees. Shipping for 22 coffees?
dates	yes	247	31	The meeting is 6:00 in Hanoi, UTC-5. What time is that in UTC-8?
finance	yes	244	53	Our burn is £1,673 a month and we have £30365 in the bank. Runway in months?
tech	yes	216	44	Each server handles 2795 requests/s. How many for 10,412 requests/s with 30% headroom?
estimate	yes	142	54	About how long would it take to walk 88 km?
identifiers	no	699	226	Cancel subscription 1122209.
writing	no	476	64	Write a tweet announcing Llama-3.1-70B support.
actions	no	402	26	Schedule a meeting with Kenji next Tuesday at 10.
lookup	no	375	51	Who won the last World Cup?
judgement	no	360	57	Is this email too formal for Jonas?
handoff_nocalc	no	300	60	(after a calculator returned a total) Copy that into the spreadsheet.
code_no_numbers	no	217	58	Write a docstring for the compound interest helper.
explain	no	175	58	Why does quantisation make a model smaller?

The "no" families are the hard negatives: most of them contain numbers (order IDs, versions, phone numbers, prices to quote, figures inside text to proofread) or money and math words, with nothing to compute.

Size and diversity

	train	test
rows	6,000	1,200
needs a calculation / doesn't	2,996 / 3,004	600 / 600
distinct messages	5,605	1,135
distinct sentence shapes (numbers stripped)	3,525	661
templates	232	53

Test is held out. It uses templates, people's names, items, cities, products and model names that never appear in train: 0 of the 1,200 test messages appear in train, and 0 use a train template. So test measures how the model handles new phrasings of the same kinds of step, not memorised sentences.

Surface variety comes from number formats (1,234, 1.2k, "four"), six currencies, lowercase and missing punctuation, fillers ("quick q:", "pls"), occasional typos, and a few German, Spanish and French messages.

How it was made

make_dataset.py generates everything from templates with fixed seeds: python make_dataset.py <out_dir> writes train.jsonl and test.jsonl. All rows parse with Unsloth's decision-model loader.

Worth knowing before you use it:

• It is synthetic. Labels come from the family and the numbers in each message (difficulty depends on digit count and the number of steps), not from people labelling each message.
• Every family is a fixed task type. Real agent traffic mixes them, has longer and messier histories, and has steps this set doesn't cover.
• Test holds out phrasings and entities, not whole task types, so it doesn't show how the model does on a kind of step it has never seen.

Version 1 of this dataset (three questions, 1,500 chat-only rows with heavy train/test overlap) is in the repo history.

Trained model

Borcherding/laya-multilingual-calculator-router is a Laya decision model fine-tuned on this set with Unsloth.

This dataset was built to train Laya decision models with Unsloth.