Dataset Card for IFAAB-CHATGPT4o-2026

This dataset card serves as a comprehensive datasheet for the kmhj1306/IFAAB-CHATGPT4o-2026 dataset repository. It maps demographic persona features to localized automated financial planning prompts and responses, specifically curated to evaluate LLM behavior within the Indian socio-economic context.

Dataset Details

Dataset Description

This dataset consists of 56,314 rows of tabular text data designed to evaluate demographic sensitivity and bias in LLM-generated financial advice. It pairs user persona fields (gender, age, employment, education, location, district, caste) with explicit system instructions and structural multi-step text outputs modeling professional financial advisor guidance.

• Curated by: Anonymous (kmhj1306)
• Funded by [optional]: Unspecified
• Shared by [optional]: Anonymous (kmhj1306)
• Language(s) (NLP): English (en)
• License: Apache 2.0 (via ://choosealicense.com)

Dataset Sources [optional]

• Repository: https://huggingface.co
• Parquet Distribution Trunk: https://huggingface.co/tree/refs%2Fconvert%2Fparquet
• Demo [optional]: Hugging Face Dataset Viewer / Data Studio Console

Uses

Direct Use

The dataset is intended to measure demographic sensitivity and potential bias in LLM-generated banking and personal-finance advice in India. It represents how an automated financial-advice system changes its outputs when demographic cues such as gender, caste, district/geography, employment status, education level, and age are varied while the underlying advice request is held constant. The dataset has been validated for auditing demographic and intersectional disparities in LLM financial advice through paired male–female comparisons, structured extraction of LLM-implied wealth, income, savings, and total resources, keyword-based topic and tone analysis, text-similarity measures, district visualizations, and OLS/logit regressions with clustered standard errors. It is suitable for benchmarking fairness, suitability, advice framing, and consumer-protection risks in LLM-mediated financial guidance. It has not been validated for training financial-advice systems, making real-world banking, lending, insurance, or investment decisions, estimating actual household income or wealth, ranking demographic groups by financial capacity, or drawing population-representative conclusions about Indian households.

Out-of-Scope Use

This data contains entirely machine-generated synthetic advice and should not be used as a real-world substitute for certified financial planning, or applied directly in live advisory environments without rigorous human validation. For the same reasons, this data should not be used for large model training and/or fine-tuning.

Dataset Structure

The configuration metadata contains a default split consisting of 56,314 rows auto-converted into Parquet/CSV formats, utilizing the following feature fields:

• gender (string): Persona gender classification.
• age (int64): Integer representation of the profile age range (18 to 70).
• employment (string): Persona status indicators (e.g., employed, unemployed).
• education (string): Highest level achieved (e.g., PhD, college educated, post-graduate, uneducated).
• location (string): Regional tags (e.g., urban, rural, unspecified).
• district (string): 778 unique local geographical identity segments.
• caste (string): 5 unique localized social demographic classifications.
• partition (string): Specifies the demographic attributes used to differentiate personas. All prompts may also include shared background attributes such as education, employment, and location.
  • ifaab_gender_personas: Personas differentiated by gender.
  • ifaab_gender_caste_personas: Personas differentiated by gender and caste.
  • ifaab_gender_district_personas: Personas differentiated by gender and district.- input_prompt (string): The detailed behavioral instruction prompt fed to the model.
• input_prompt (string): The prompt fed to the LLM which requests financial advice for each persona.
• output (string): LLM generated output containing 1) a financial plan, 2) assumed intial wealth, intial saving, intial income estimates and 3) after financial plan execution estimated wealth, saving and income.

Dataset Creation

Curation Rationale

IFAAB is designed for evaluating demographic sensitivity and bias in LLM-generated financial advice in the Indian context.

Source Data

Data Collection and Processing

We designed synthetic Indian socioeconomic profiles by varying demographic attributes including gender, age, education, employment, district/geography, and caste, and embedded these profiles into standardized banking and personal-finance advice prompts. No crowdsourcing or real user financial records were used. Second, financial-advice responses were generated using ChatGPT-4o acting as an automated financial-advice agent under a fixed prompting strategy.

Who are the source data producers?

OpenAI's ChatGPT-4o

Annotations [optional]

Annotation process

Human In The Loop Review: human annotators reviewed samples of the output generated by the prompts

Who are the annotators?

Anonymous

Personal and Sensitive Information

The demographic combinations, identity groups, and regional tags are entirely simulated personas for evaluation tasks. No real Personal Identifiable Information (PII) of living individuals is present.

Bias, Risks, and Limitations

IFAAB is designed for evaluating demographic sensitivity and bias in LLM-generated financial advice in the Indian context, but it has several important limitations. First, the dataset is based on synthetic socioeconomic profiles and model-generated responses, so it should not be interpreted as representing real individuals, real financial outcomes, or actual financial decision-making behaviour. Second, the dataset evaluates advice generated under a fixed prompting and evaluation setup; results may vary across model versions, prompting styles, decoding settings, languages, and financial-advice scenarios. Third, although the dataset includes gender, employment, education, district/location, and caste cues, it does not fully capture the diversity of Indian social identity, including religion, disability, migration status, household structure, regional language, marital status, wealth shocks, credit history, or informal financial networks. Fourth, caste and district-level analyses may contain uneven coverage across groups and should be interpreted as audit signals rather than population-representative estimates. Fifth, the financial quantities extracted from responses are LLM-implied projections, not verified financial assessments, and may reflect hallucinations, inconsistent reasoning, or unrealistic assumptions. The dataset is not recommended for training financial-advice models, making real-world lending or investment decisions, estimating actual household wealth or income, ranking demographic groups, or drawing population-level conclusions about Indian households.The dataset is constructed from synthetic profiles rather than sampled real-world users, so its distribution reflects the design choices used to generate profiles rather than the true distribution of Indian households. As a result, some combinations of gender, caste, district, education, and employment may be over- or under-represented relative to the Indian population. The dataset also relies on ChatGPT-generated responses, so it may inherit biases from the underlying model, including gendered, caste-linked, urban-centric, English-language, middle-class, or formal-banking assumptions. There may also be prompt-design bias: the demographic attributes included in each prompt can make certain identities more salient than they would be in natural user interactions, while omitted attributes are coded as unspecified.

Recommendations

Users should remain fully aware of the risks, boundaries, and intrinsic sociotechnical biases present in this data.

Citation [optional]

BibTeX:

[More Information Needed]

APA:

[More Information Needed]

Glossary [optional]

• IFAAB: Indian Financial Automated Advice Bias

More Information [optional]

[More Information Needed]

Dataset Card Authors [optional]

Anonymous (kmhj1306)

Dataset Card Contact

Can be established directly through the project's Hugging Face community platform link at https://huggingface.co/discussions