small-models-for-glam/synthetic-parsed-names-yaml

Dataset Card for Synthetic Parsed Names (YAML)

This dataset contains approximately 500,000 synthetic examples of complex, unstructured historical names paired with their structured YAML equivalents. It is designed to fine-tune small open-source large language models (LLMs) to accurately parse cultural heritage name strings into isolated components (first names, last names, middle names, dates, titles, etc.) for de-duplication and structured data ingestion.

Dataset Details

Dataset Description

This dataset was created to solve a specific entity extraction problem in the cultural heritage sector: accurately parsing highly varied historical name strings into structured data formats. When commercial frontier models proved too expensive and structurally inconsistent (often generating broken JSON), the creators shifted to training small parameter open-source models (like Qwen 0.5B to 4B) to output valid YAML.

To train these models effectively, half a million synthetic input/output pairs were generated. The synthetic data mimics the variance in naming conventions across different time periods and geographies, accommodating localized titles, suffixes, and date qualifiers.

• Curated by: Yale University LUX Platform Team / small-models-for-glam
• Language(s) (NLP): English (en), French (fr), German (de), Italian (it), Arabic (ar), and Chinese (zh) naming conventions and qualifiers.
• License: MIT (or applicable open-source license)

Dataset Sources

• Repository: small-models-for-glam/synthetic-parsed-names-yaml
• Associated Models: - Qwen3.5-0.8B-SFT-name-parser-yaml
  • Qwen3.5-2B-SFT-name-parser-yaml
  • Qwen3.5-4B-SFT-name-parser-yaml

Uses

Direct Use

The primary use case is Supervised Fine-Tuning (SFT) of small language models for precise entity extraction tasks. Specifically, the dataset trains models to:

1. Accept an unstructured name string.
2. Return a strict, syntactically valid YAML block isolating first_name, last_name, middle_names, temporal (life spans, flourished dates), titles, and extra_info.

Out-of-Scope Use

Because the names and dates in this dataset are entirely synthetic and generated via a Python script, this dataset should not be used for historical research, biographical fact-checking, or generating factual information about real historical figures.

Dataset Structure

The dataset is formatted in a standard conversational standard (messages list) suitable for chat-based fine-tuning.

Each row contains a messages array with two dictionaries:

1. User Message: Contains the prompt (e.g., "Parse this person name:\n\nCynthia Ponce'Hayes (born 1633), born 1633").
2. Assistant Message: Contains the expected structured YAML output.

Example output structure:

first_name: Cynthia
last_name: Ponce'Hayes
middle_names: []
temporal:
- start: 1633
  end: null
  type: life_span
titles: []
extra_info: []

Dataset Creation

Curation Rationale

Yale University's LUX platform contains over 3 million individual person records. To programmatically identify duplicate identities across 3.6 million parsed historical names, the team needed reliable structured data. General-purpose frontier models had ~70% accuracy and struggled with syntax, while commercial API costs were prohibitive at scale. This synthetic dataset was curated to locally train small, efficient open-source models to perform this narrow extraction task with 94-96% accuracy.

Source Data

Data Collection and Processing

The data is 100% synthetically generated using a Python script utilizing the Faker library.

To mimic the complexities of real-world cultural heritage data, the generation script uses custom template structures. The team built a CulturalHeritageProvider class to extend Faker, introducing specific linguistic and historical variants, such as:

• English: "Duke", "Sir", "the Elder", "circa"
• French: "Comte", "le Jeune", "vers", "né", "mort"
• (Along with rules for German, Italian, Arabic, and Chinese locales).

This ensures the generated input strings effectively mimic the messy, localized data found in archives and museum catalogs.

Who are the source data producers?

The data was procedurally generated by the LUX platform team at Yale University.

Personal and Sensitive Information

This dataset contains no personal, sensitive, or private information. All names, dates, titles, and combinations thereof were procedurally generated using randomized algorithms. Any resemblance to real historical or living persons is purely coincidental.

Bias, Risks, and Limitations

While the custom CulturalHeritageProvider simulates complex historical data, it may not encapsulate the entirety of edge cases present in 3.6 million real-world, unstructured archival records.

Furthermore, synthetic data can occasionally generate minor systemic anomalies. For instance, models trained on similar synthetic patterns might occasionally misclassify or misspell certain qualifiers (e.g., predicting "floured" instead of "flourished" if not properly represented or if a typo existed in the generation template).

Recommendations

Users should be aware that the dataset is highly tailored for cultural heritage name structures (GLAM sector). If fine-tuning models for modern, standard name processing (e.g., standard CRM databases without historical titles or ambiguous life dates), simpler datasets may suffice.

Glossary

• GLAM: Galleries, Libraries, Archives, and Museums.
• LUX: Yale University's cultural heritage collections discovery platform.
• Temporal Types: Classifications for historical dates (e.g., life_span, flourished, active).