Uliana333/ukrainian-refugees-financial-advisory

Ukrainian Refugees Financial Advisory Dataset

A dataset of 500 synthetic advisory cases generated by a multi-agent LLM pipeline that produces and evaluates retirement-oriented financial guidance for Ukrainian refugee-like profiles in Poland.

Each case covers one full advisory cycle: synthetic profile generation → draft recommendation + clarifying questions → final structured recommendation → automated quality evaluation.

GitHub: uliana0203/ai-agents-refugee-finance

Dataset summary

Metric	Value
Total cases	500
Accepted	327 (65.4%)
Needs review	170 (34.0%)
Flagged	3 (0.6%)
Mean composite score	8.64 / 10
File size	~17.6 MB

Motivation

Ukrainian refugees in Poland face specific financial challenges: navigating an unfamiliar pension system, managing tight household budgets, and making long-term retirement decisions under uncertainty. This dataset was created to study whether a RAG-grounded multi-agent pipeline can produce safe, actionable, and well-sourced financial guidance for this population - and to benchmark the quality of LLM-generated financial advice against deterministic and rubric-based evaluation criteria.

Synthetic profiles are calibrated to the 2025 National Bank of Poland survey of Ukrainian refugees (N = 3,965).

Dataset structure

Each record is a JSON object with the following top-level fields:

Field	Type	Description
run_id	string	UUID for the advisory case
ts_unix	int	Unix timestamp of generation
case_status	string	accepted, needs_review, or flagged
refugee	object	Synthetic client profile
consultant_draft	object	Draft recommendation + clarifying questions
refugee_clarifying_qa	list	Clarifying Q&A exchange
consultant_final	object	Final structured recommendation
evaluator	object	Deterministic checks + LLM rubric scores

refugee

Field	Description
anchors	Demographic inputs: gender, dependents, employment status, months in Poland, income range, remittances, language level
constraints	4 qualitative constraints grounded in refugee-context RAG
persona_json	LLM-generated persona: age, education, occupation, financial literacy, risk tolerance, retirement goals, etc.
profile_json	Budget simulation output: income, savings, housing, food, transport, healthcare, remittances, surplus
profile_text	Plain-text profile summary passed to the Consultant agent
user_query	Retirement-oriented user question

consultant_draft / consultant_final

Draft contains the preliminary answer and up to 3 clarifying questions. Final contains the structured recommendation with mandatory sections: summary, budget check, suggested monthly saving amount, retirement options in Poland, next steps, and sources.

evaluator

Field	Description
deterministic_checks	Citation presence, source validity, arithmetic consistency (±PLN 30), required sections, private-program mentions
llm_rubric.scores	Five dimensions scored 0–10: groundedness, arithmetic consistency, actionability, clarity, safety/ethics
llm_rubric.issues	Detected issue types: missing_citation, unsupported_claim, unsafe_overconfidence, math_error, private_program_hallucination
final.overall_score	Composite score (mean of rubric dimensions + deterministic penalties, clamped 0–10)
final.allow_to_show	Whether the recommendation passed the showability check

Usage

from datasets import load_dataset

ds = load_dataset(
    "Uliana333/ukrainian-refugees-financial-advisory",
    data_files="data/advisory_cases.jsonl",
    split="train",
)

# Filter to accepted cases only
accepted = ds.filter(lambda x: x["case_status"] == "accepted")

# Access a recommendation
print(accepted[0]["consultant_final"]["final_answer"])

# Access evaluator scores
print(accepted[0]["evaluator"]["llm_rubric"]["scores"])