Hammer-Purgstall Correspondence TEI Evaluation Dataset

This dataset supports LLM-supported TEI encoding tasks, output evaluation and comparison. It comprises a representative sample of 100 letters from the correspondence of Joseph von Hammer-Purgstall (1774-1856), an Austrian orientalist, historian, and diplomat. The correspondence spans six decades (1790s-1850s) and exhibits significant linguistic diversity, containing letters primarily in German alongside English, French, and Italian, with instances of code-switching (Latin, Greek, and Arabic text segments). The dataset includes four main components for LLM processing:1. Input component: 100 plain text letter transcriptions (UTF-8)2. Reference component: 100 manually encoded TEI XML (P5) reference annotations following TEI Guidelines for correspondence3. Output component: LLM-generated TEI XML encodings from four models (GPT-5-mini, Claude Sonnet 4.5, Qwen3-14B-Q6, OLMo2-32B-instruct-Q4), totaling 400 encodings4. Evaluation component: Comprehensive assessment results in JSON format with aggregate Excel reports and a visualization of the cross-model comparison The sample was selected through systematic stratified sampling based on language distribution, writer diversity, and letter length variation. Files included:- hpe-correspondence-metadata.json: Metadata for 100 letters (language, sender, recipient, date, letter length, etc.)- hpe-correspondence-transcriptions.zip: Plain text letter transcriptions- hpe-correspondence-tei-reference.zip: Manually encoded TEI XML reference files- hpe-correspondence-llm-encodings.zip: LLM-generated TEI XML encodings from four models- hpe-correspondence-evaluation.zip: Evaluation results in JSON format with Excel reports- hpe-prompt-templates.zip: Five prompt scenarios for LLM processing with encoding instructions and few-shot samples Code Availability The evaluation results in this dataset were generated using the TEI LLM Evaluation Framework. The source code is available at: https://github.com/strubrina/tei-evaluation.git