Fine-tuning tutorial: train ModelSpanExtractor on a custom domain

[adaamko]

Summary

Write a tutorial (notebook or markdown guide) showing how to fine-tune ModelSpanExtractor on custom domain data using the existing training pipeline in verbatim_rag/extractor_models/.

Motivation

The verbatim-rag-modern-bert-v2 model was trained on scientific papers, medical literature, financial tables, and other domains, but users in niche domains (legal contracts, internal documentation, code output) may want a specialized extractor. The training code exists but is undocumented for external users.

Scope

1. Data format: how to structure {question, document, gold_spans[]} pairs as a training dataset
2. Training: python verbatim_rag/extractor_models/train.py --data_path data/ --output_dir output/
3. Evaluation: word-F1 on a held-out dev set
4. Loading the fine-tuned model: ModelSpanExtractor(model_path="./output/checkpoint-best")

A Jupyter notebook in examples/ or a docs/fine_tuning.md are both acceptable.

Notes

• The training script CLI already exists at verbatim_rag/extractor_models/train.py
• QADataset, QAModel, and Trainer are in verbatim_rag/extractor_models/
• A small synthetic dataset example would make the tutorial self-contained