Fine-Tuned Models

• mT5-small (300M parameters)
• GPT-2 Small (124M parameters)
• LLaMA 3.1 (1B parameters)

Current Training Data

• 100 pages: 40,000 lines (~0.4M words)
• 200 pages: 80,000 lines (~0.8M words)
• 400 pages: 160,000 lines (~1.6M words)

Objectives

• Word Correction: Identify and correct errors using contextual and grammatical cues.
• Language Identification: Distinguish English from Latin text.
• Artefact Removal: Eliminate HTR-generated artefacts.
• Structural Recognition: Detect depositions’ components (e.g., front matter, headings, articles).
• Insertion Logic: Handle missing text at marked positions.

1.2 Integration with RAG Pipeline

Components:

• Retriever: BM25 or Sentence-BERT
• LLM: mT5-small
• Corpus: Curated historical texts or JSON/SQLite databases

Deployment Highlights:

• Scalable: Easily runs on platforms like Hugging Face Spaces with lightweight GPU instances.
• API-Friendly: Supports integrations via Hugging Face Inference API for retrieval-augmented tasks.

📚 2.0 Datasets

2.1 Published Datasets

ENGLISH HIGH COURT OF ADMIRALTY DEPOSITIONS

1. MarineLives/English-Expansions
2. MarineLives/Latin-Expansions
3. MarineLives/Line-Insertions
4. MarineLives/HCA-1358-Errors-In-Phrases
5. MarineLives/HCA-13-58-TEXT

YIDDISH LETTERS

1. MarineLives/Gavin-yiddish-raw-HTR-and-groundtruth-lines
2. MarineLives/Gavin-yiddish-raw-HTR-and-groundtruth-paragraphs

2.2 Unpublished Datasets

• Dataset 1: 420K tokens, full diplomatic transcription (1627–1660)
• Dataset 2: 4.5M tokens, semi-diplomatic transcription (1607–1660)
• Dataset 3: 100K tokens, diplomatic transcription of Early Modern letters (1600–1685)

🌍 Explore MarineLives

Join us in unlocking Early Modern history by exploring our Hugging Face organization and datasets! You can follow us on BlueSky at @marinelives.bsky.social You can explore our content on our MarineLives wiki and on our ai-and-history-collaboratory GitHub repository.