The problem
Seventeenth-century handwritten records contain irregular letterforms, unfamiliar spellings, and lines that can admit several plausible readings. Turning cropped archival images into text calls for more than applying an off-the-shelf OCR model. It requires adapting visual readers, examining their disagreements, and preserving enough evidence to explain which experiment produced a transcription.
This project brings two phases of work on Barbados archival handwriting into one research effort: an initial model-development pipeline and a later experiment system focused on traceability and reproduction.
From line images to candidate transcripts
The first phase organizes training and inference around configurable vision-language models. LoRA adaptation makes it possible to explore model changes without treating every run as a full fine-tune. Image preparation, generation settings, and evaluation live in shared Python modules, with notebooks acting as launchers for remote GPU experiments.
Validation folds group near-duplicate text so closely related examples do not casually cross the training boundary. Candidate experiments use a development fold, with a separate confirmation fold supporting decisions about whether to advance a recipe.
Multiple readers produce alternative transcriptions for each line. Minimum-Bayes-risk selection compares those candidates under an edit-based loss and chooses a transcript with low expected disagreement. Word voting and guarded character-level consensus provide additional approaches to investigate. Their usefulness is checked on held-out references instead of assumed from the number of models involved.
Making experiments traceable
As the experiment collection grew, I separated the system into a local controller, stage notebooks, and a reusable research library. The controller transports committed source and manages remote jobs. Notebooks expose each experiment stage. The library contains the actual preparation, training, evaluation, and ensemble logic.
Typed configuration records the exact recipe. Pinned data and model references, source fingerprints, locked dependencies, and artifact manifests connect a result to the inputs that produced it. Model bundles can be reloaded for independent validation, and replay commands compare fresh outputs with recorded artifacts.
That structure also keeps test prediction separate from evaluation: an unlabeled test run writes ordered transcriptions without presenting a test accuracy score.
Results and research lessons
The work produced an extensible transcription pipeline and a repeatable way to compare visual readers and ensembles. It also preserved unsuccessful experiments. Recorded augmentation trials and candidate-judging approaches did not meet their advancement criteria, so those ideas were retained as negative findings rather than silently becoming the preferred method.
Later experiments explored whether smaller reader subsets could reproduce a larger ensemble. Agreement with an archived transcription is useful for that compression question, but it does not establish correctness against human labels. Keeping those objectives separate is a central part of this ongoing research.