williamjmorenor/personal-finance-chatml-dataset

Bilingual Personal Finance ChatML Dataset (EN/ES)

Dataset Description

This dataset is a professionally curated bilingual (English/Spanish) instruction dataset designed for fine-tuning large language models (LLMs) in the domain of personal finance.

It is structured in ChatML format and intended for supervised fine-tuning (SFT), domain adaptation, and financial instruction modeling.

The dataset is created and reviewed from an accounting perspective, ensuring conceptual precision and terminological consistency.

Supported Languages

• —English (en)
• —Spanish (es)

Each record includes a language field for filtering and evaluation.

Dataset Structure

Each entry follows this structure:

{
  "pair_id": "000001",
  "language": "en | es",
  "domain": "personal_finance",
  "topic": "string",
  "level": "basic | intermediate | advanced",
  "system": "...",
  "question": "...",
  "answer": "...",
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Fields

• —pair_id: Alignment identifier shared by equivalent EN/ES rows.
• —language: Language of the interaction.
• —domain: Domain classification (for this dataset: personal_finance).
• —topic: Thematic classification.
• —level: Conceptual difficulty level.
• —system: System instruction prompt.
• —question: User question in flat format (equivalent to messages[1].content).
• —answer: Assistant answer in flat format (equivalent to messages[2].content).
• —messages: ChatML conversation format (system, user, assistant).

Topics Covered

• —Foundations of personal finance
• —Budgeting
• —Income and expense management
• —Time value of money
• —Compound interest
• —Debt management
• —Rollforward analysis
• —Net worth tracking
• —Savings strategies
• —Investment basics
• —Risk management
• —Retirement planning
• —Topics may expand in future versions.

Annotation Process

All examples are manually curated and written to ensure:

• —Technical accounting accuracy
• —Terminological consistency
• —Pedagogical clarity
• —Neutral regulatory positioning

The dataset avoids country-specific legal advice unless explicitly indicated.

Intended Uses

This dataset is intended for:

• —Supervised Fine-Tuning (SFT)
• —Instruction tuning
• —Financial domain specialization
• —Educational financial assistants
• —Research on bilingual financial NLP

Out-of-Scope Uses

• —Automated financial advisory systems without human oversight
• —Regulatory or tax compliance automation
• —High-stakes investment decision systems

This dataset is educational and instructional in nature.

Data Splits

If provided:

• —Train
• —Validation
• —Test

Otherwise, users may split according to their experimental needs.

Current Dataset Snapshot (2026-03-08)

• —Raw base source (https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/raw/personal_finance_dataset.csv): 3,781 lines (including header).
• —Raw seed source (https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/raw/data_sed.csv): 1 line (header only at this snapshot).
• —Processed full ChatML (personal_finance_chatml_full.jsonl): 17,580 records.
• —Processed messages-only (data/processed/personal_finance_chatml_messages.jsonl): 17,580 records.

Note: The processed files are generated through the amplification pipeline and therefore contain the expanded bilingual training set.

Data Amplification Techniques Implemented

The project implements deterministic template-based semantic amplification in https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/scripts/dataset_amplification.py with two complementary strategies:

1. 1.Seed semantic amplification (from https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/raw/data_sed.csv)
2. 2.Uses three bilingual seed template files:
3. 3.https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/amplification_templates/seed_definition_templates.json
4. 4.https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/amplification_templates/seed_variants_templates.json
5. 5.https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/amplification_templates/seed_examples_templates.json
6. 6.Expands each seed concept into multiple EN/ES question variants via cadenas_amplificadas.
7. 7.Uses field-controlled answers (definition_*, variants_*, examples_*) to preserve intent consistency.

1. 1.Base pattern amplification (from https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/raw/personal_finance_dataset.csv)
2. 2.Uses paired ES/EN pattern files:
3. 3.https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/amplification_templates/base_rephrase_templates.json
4. 4.https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/amplification_templates/base_noisy_templates.json
5. 5.Detects original user prompts with regex_cadena_original and extracts {concept} using named groups.
6. 6.Generates bilingual rephrased and noisy variants while preserving original assistant answers and metadata (system, topic, level, domain).

1. 1.Alignment and quality controls during amplification
2. 2.Bilingual pair enforcement (ES/EN counterpart generation per synthetic pair).
3. 3.Template validation (required sections/fields and non-empty amplified lists).
4. 4.Duplicate prevention using signature checks (language, user, assistant).
5. 5.Dynamic pair_id regeneration in output generation (https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/scripts/csv_to_jsonl.py) starting at 000001.

Current note: Seed amplification logic is implemented and active in the pipeline, but the current https://github.com/williamjmorenor/personal-finance-chatml-dataset/blob/main/data/raw/data_sed.csv snapshot contains only headers, so current growth is effectively driven by base pattern templates.

Bias, Risks, and Limitations

• —The dataset reflects a structured accounting perspective.
• —It does not include regional tax laws or jurisdiction-specific regulations by default.
• —It does not provide personalized financial advice.
• —It may not reflect emerging financial instruments or regulatory changes.

Versioning

• —v2.0 — Amplified dataset

• —v1.0 — Initial bilingual release
• —Future versions may expand:
• —Advanced investment topics
• —Behavioral finance

Citation

If you use this dataset in research, please cite:

@dataset{bilingual_personal_finance_chatml,
  author = {William José Moreno Reyes (CP/MBA)},
  title = {Bilingual Personal Finance ChatML Dataset (EN/ES)},
  year = {2026},
  publisher = {BMO Soluciones, S.A.}
}

License

MIT License (or chosen license).

Contact

For questions, suggestions, or collaboration opportunities, please open an issue in the repository.