mghafiri/decision-model-scenarios

Decision Model Scenarios

2000 synthetic English scenarios with 9716 typed questions and calibrated soft labels. It is the training data for mghafiri/qwen3.5-0.8B-decision-model, a small model that answers typed questions about a text "state" with probability distributions instead of generated text.

Each scenario has three parts:

• a state: a message, a record, an email thread, a policy plus a ticket, a log, and so on
• 3–6 independent questions of three types:
  • Choice: pick one option
  • Score: pick a level on an ordered scale
  • Noul: yes/no
• a target probability distribution for each question, with a difficulty tag and a short reasoning note

The question format follows the publicly documented request shape of TypeSafe's System One API (docs.typesafe.ai). This dataset is independent and not affiliated with TypeSafe.

Splits

split	scenarios	questions
train	1700	8282
validation	150	720
test	150	714

• Domains (32): agent_tool_routing, airline_travel_ops, aml_kyc_compliance, app_marketplace_policy, banking_fintech, citation_verification, clinical_literature_screening, customer_support, cybersecurity_soc, ecommerce, education_admissions, energy_utilities, entity_matching, government_services, healthcare_admin, hr_recruiting, insurance_claims, it_helpdesk, legal_compliance, llm_guardrails, logistics_sales_crm, manufacturing_qa, news_claim_verification, payroll_benefits, procurement_vendor_risk, property_management, rag_passage_relevance, smart_home_iot, software_engineering, span_selection_extraction, telecom_support, trust_safety_moderation.
• Question types: noul 41%, choice 37%, score 22%.
• Difficulty: clear 39%, moderate 31%, borderline 16%, insufficient 9%, adversarial 6%.

About half of the scenarios are realistic multi-document cases where the correct label needs multi-step reasoning: policy exceptions, later corrections, 2-hop lookups, and instructions planted in the text that must be ignored. Each question is still a single literal judgment.

Fields

field	type	content
id	string	scenario id (bNN-NNNN)
domain	string	one of the 32 domains
pattern	string	fan-out, routing, detection, scoring, verification, extraction-choice, ranking-relevance, matching, guardrail
state	JSON string	the content to judge (string, object or array)
questions	JSON string	`{question_id: {"type": "choice"
targets	JSON string	{question_id: {"probabilities": {...}} or {"noul": p}, "difficulty": ..., "note": ...}}

• Choice targets are keyed by option.
• Score targets are keyed by level index ("0"…"K-1").
• Noul targets give noul = P(yes).

import json
from datasets import load_dataset

ds = load_dataset("mghafiri/decision-model-scenarios")
row = ds["train"][0]
state, questions, targets = (json.loads(row[k]) for k in ("state", "questions", "targets"))

How the labels were made

1. Authoring. Claude (Anthropic) wrote each scenario and its targets. It followed a written rubric (DATA_SPEC.md): the target probability of an option is the fraction of careful domain experts who would pick it.
2. Blind second annotation. An independent Claude annotator relabelled every question without seeing the original labels. Agreement on the top answer was 9434/9718 = 97.1%.
3. Merge:
   • Where the two agreed, the target is 0.6·author + 0.4·reviewer.
   • Borderline disagreements were averaged 50/50.
   • The remaining disagreements were adjudicated one by one, and defective questions were removed.
4. Automated checks:
   • schema and probability sums
   • near-duplicate states
   • synthetic-only personal data (example.com emails, 555-01xx phones, fictional organisations)
   • no arithmetic, counting or date-comparison questions

Minimum achievable log-loss on these soft targets: about 0.36, the mean target entropy.

Limitations

• The data is synthetic and LLM-authored. The targets estimate expert agreement. They are not real-world outcomes.
• English only. The domains are business and operations workflows; there is no clinical or legal advice content.
• Personal data is synthetic. Any resemblance to real people or organisations is unintended.

License

The data is released under the MIT License; see LICENSE.