README.md

metadata

base_model: Qwen/Qwen3-0.6B
library_name: transformers
model_name: Qwen3-0.6B-SFT-name-parser-yaml
tags:
  - generated_from_trainer
  - trl
  - sft
  - name-parsing
  - cultural-heritage
  - yaml
  - nlp
license: apache-2.0
language:
  - en
  - multilingual
pipeline_tag: text-generation

Model Card for Qwen3-0.6B-SFT-name-parser-yaml

This model is a fine-tuned version of Qwen/Qwen3-0.6B specifically designed for parsing cultural heritage person names into structured YAML format. It has been trained using TRL with supervised fine-tuning (SFT).

Model Description

This specialized model parses person names from cultural heritage contexts (libraries, archives, museums) into structured YAML format with the following fields:

• first_name: Person's given name
• last_name: Person's family name or surname
• middle_names: List of middle names or initials
• temporal: List of temporal information (birth, death, flourished dates)
• titles: List of titles, honorifics, or professional designations
• extra_info: List of additional information (places, affiliations)

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "small-models-for-glam/Qwen3-0.6B-SFT-name-parser-yaml"

# Load the tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Parse a person name
input_name = "Dr. Jane Smith-Jones, 1850-1920"
prompt = "Parse this person name:\n\n" + input_name

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(**model_inputs, max_new_tokens=512)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# Parse thinking content if present
try:
    index = len(output_ids) - output_ids[::-1].index(151668)  # </think> token
except ValueError:
    index = 0

content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip()
print(content)

Expected output:

first_name: Jane
last_name: Smith-Jones
middle_names: []
temporal:
- start: 1850
  end: 1920
  type: life_span
titles:
- Dr.
extra_info: []

Supported Name Patterns

The model handles a wide variety of name formats commonly found in cultural heritage contexts:

Basic Patterns

• John Smith
• Smith, John
• Dr. John Smith
• John A. Smith

Complex Patterns

• Baron William Henry Ashe A'Court Heytesbury, c. 1809-1891
• Jones, James Earl, Dr., (fl. 1850-1900)
• Miller, Chester F. (Chester Frederic), 1886-
• Rábade Obradó, Ana Isabel
• 彭大铨 (Chinese names)

Edge Cases

• Mononyms: Salzmann, Mokamba
• Initials: J. F. Vitry, A. E. Borie
• Diacritics: Péporté, Gerencsér
• Temporal data: Rosana, 1963-
• Parenthetical expansions: T. (Takeshi) Ohba

Training Procedure

Training Data

The model was trained on a synthetic dataset of 1,000+ examples generated using a comprehensive template-based approach that covers:

• 70% regular examples: Standard name patterns with various combinations of fields
• 30% edge cases: Challenging patterns including mononyms, initials, diacritics, and non-Western names

Data Generation Features

• Multi-cultural support: Names from English, French, German, Italian, Spanish, Dutch, Arabic, and Chinese contexts
• Temporal data variety: Birth/death dates, flourished periods, single dates
• Title diversity: Academic, religious, nobility, military, and professional titles
• Complex surnames: Hyphenated, apostrophized, and particle-based surnames (van, von, de, al-, ibn)

Training Configuration

• Base model: Qwen/Qwen3-0.6B
• Training method: Supervised Fine-Tuning (SFT) using TRL
• Output format: YAML with consistent field ordering
• Chat template: Standard user/assistant format with "Parse this person name:" prompt

Framework Versions

• TRL: 0.23.0
• Transformers: 4.56.2
• PyTorch: 2.8.0
• Datasets: 4.1.1
• Tokenizers: 0.22.1

Performance

The model demonstrates strong performance on cultural heritage name parsing tasks:

• Handles diverse international name formats
• Correctly identifies and structures temporal information
• Processes titles, honorifics, and professional designations
• Manages complex surname patterns and particles
• Supports mononyms and abbreviated names

Limitations

• Primarily trained on Western and East Asian name patterns
• May struggle with very rare or highly specialized naming conventions
• Temporal date parsing assumes Gregorian calendar years
• Limited support for ancient or historical dating systems (BCE, regnal years)

Intended Use

Primary Use Cases

• Digital humanities: Processing historical person names in manuscripts and documents
• Library science: Cataloging and standardizing author names in bibliographic records
• Archive management: Structuring person names in archival finding aids
• Museum collections: Organizing creator and subject names in cultural heritage databases

Out-of-Scope Use

• Modern person name parsing for contemporary applications
• Legal document processing requiring high precision
• Real-time person identification or verification
• Processing of fictional character names

Ethical Considerations

• The model reflects naming conventions present in its training data
• Cultural biases may exist toward Western naming patterns
• Should not be used for identity verification or legal purposes
• Consider cultural sensitivity when processing names from different traditions

Framework Citation

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}

Model Card Contact

For questions about this model card or the model itself, please open an issue in the project repository.