Brainquiver/classify-typed-general-finetune-preview

Classification · Typed Decisions · General · Finetuning · Preview (Synthetic)

This preview dataset holds 20,000 samples of typed decisions from 20 business workflows. Each sample pairs a state, such as an invoice, a support thread or a contract clause, with one to eight questions about it. A question can be a choice, a score or a yes or no, and its gold gives a probability for every option. The full private dataset holds 8,006,319 samples of 105 workflows.

Composition

The preview has one config, default, with one split, train. The samples come in three equal shares.

Source	Samples	Workflows	Text
setmill	6,667	8	composed by setmill alone
Llama 3.3 70B Instruct	6,667	12	composed by setmill, prose written by Llama 3.3 70B Instruct
Llama 4 Maverick	6,666	12	composed by setmill, prose written by Llama 4 Maverick

The two LLM writer shares hold the same 12 workflows, and each writer workflow has 1,110 to 1,112 samples, about half from each model. The 8 setmill workflows have 833 or 834 samples each. The kind of a sample says how much of its state a model wrote.

Kind	Samples	State
structured	6,667	a JSON record, with every value composed
mixed	9,943	a JSON record with written parts, such as a message or a description
prose	3,390	one written text, such as an email or a post

Format

Each record has seven fields. state, questions and gold are JSON strings, as in the full set.

Field	Holds
id	the workflow and a number of five digits, such as ticket-triage-00001
workflow, area	the workflow and its area
state	the record that the questions are about
questions	each question with its type, its instructions and its options
gold	the answer to each question, with a probability for every option
n_questions	the number of questions

The preview leaves out the full set's metadata, so it has no factors, briefs or writer names. The questions come in three types as showcased in the table below.

Type	Questions	%	Options	Gold keys
noul	44,485	52.79	a yes or no, with an optional description of each answer	false and true
choice	28,230	33.50	2 to 20 named options	the option keys
score	11,553	13.71	2 to 10 levels, lowest first	"0" to "n-1"

A sample asks 4.21 questions on average. In gold, probabilities covers every option shown and sums to 1, and label is the most probable option. A clear case puts 1.0 on its answer, while a case that the rules leave open spreads the probability across its readings. The other gold fields follow from the probabilities. confidence is (n x p_max - 1) / (n - 1) for a choice of n options and 2 x p_max - 1 for a yes or no. A score also adds score, the expected level, and a yes or no adds noul, the probability of true.

In 18 of the 20 workflows, every gold is a clear case. The other two keep the soft golds of their tone question at the share that the full set gives them.

Workflow	Question	Samples with a soft gold	Top probability
ticket-triage	tone	588 of 1,110 (53.0%)	0.6 or 0.8
chargeback-review	cardholder_tone	619 of 1,111 (55.7%)	0.7 or 0.8

Workflows

Workflow	Area	Source	Samples
invoice-po-match	accounts-payable	setmill	833
tool-call-approval	ai-agents	both writers	1,110
loan-screening	banking-and-lending	setmill	833
chargeback-review	billing-and-payments	both writers	1,111
transaction-monitoring	compliance-and-risk	setmill	833
ticket-triage	customer-support	both writers	1,110
extenuating-circumstances	education-administration	both writers	1,112
patient-message	healthcare-administration	both writers	1,111
candidate-screening	hr-and-recruiting	setmill	834
home-claim-cover	insurance-claims	both writers	1,112
incident-triage	it-operations	setmill	834
contract-review	legal-and-contracts	both writers	1,112
shipment-exception	logistics-and-shipping	setmill	833
ad-claim-review	marketing-and-content	both writers	1,111
maintenance-request	real-estate-and-property	both writers	1,111
return-request	retail-and-ecommerce	both writers	1,110
renewal-risk	sales-and-crm	setmill	833
alert-triage	security-operations	setmill	834
bug-report-triage	software-engineering	both writers	1,111
content-moderation	trust-and-safety	both writers	1,112

Construction

Our composer built every sample while our LLM generator wrote each part. A check then tested the text against the brief. The preview takes its samples from the curated full set, where inside each workflow, each layout keeps its share of the full set. The labels of the workflow's first question stay as balanced as the workflow allows.

Verification

Each candidate passed the full set's check of every rule, and the full set's clean-up processes. A read then covered 835 samples across multiple stages, each read in full against its workflow's rules and its gold.

Usage

import json
from datasets import load_dataset

ds = load_dataset("Brainquiver/classify-typed-general-finetune-preview")

row = ds["train"][0]
state = json.loads(row["state"])
questions = json.loads(row["questions"])
gold = json.loads(row["gold"])

# The preview has only a train split. Hold out whole workflows for a test of new decisions.
held = {"contract-review", "renewal-risk"}
train = ds["train"].filter(lambda r: r["workflow"] not in held)
test = ds["train"].filter(lambda r: r["workflow"] in held)

Intended Use

This preview gives a first look at the full set, and suits finetuning or evaluating a model that answers typed questions about a record. It has only a train split, so a user splits it at load time. A test of how a model carries over to new decisions therefore holds out whole workflows. By comparison, a random split of samples measures recall of the workflows it trained on.

Limitations

The data is synthetic and English only. The records are composed for the set, but their names, cities, countries and companies come from shared tables that mix real and invented entries. A drawn value, such as a phone number or an amount, can also match a real one by chance. A workflow's rules are invented too, as a plausible policy written for the set.

The preview holds 20 of the 105 workflows, one for each area, so it shows a narrow part of each area. Due to the synthetic model-generated nature of the samples, some known faults can remain in the preview, where patterns cannot find them.

Class	Effect
prose that contradicts its own record in a way that no pattern finds	the golds follow the composed record
question thresholds that differ from the shown policy	each question states its own threshold, and its gold follows it
option subsets, where a choice shows some of its policy's outcomes	the gold option is always shown

The prose comes from two models and carries their habits, such as a narrow range of openings, a few favorite phrases and a formal register.

License and Provenance

The workflows, the composer, the composed records, the questions, the golds and the curation are the work of Brainquiver. Brainquiver releases this work under CC BY-NC 4.0. The prose of the writer samples, however, is the output of Llama 3.3 70B Instruct and Llama 4 Maverick, under those models' community licenses. The dataset's license is therefore marked as other.