MonumentalSystems/text-pipeline-capitalization-lora

Text Pipeline Capitalization Adapter

This repository preserves the capitalization adapter and intermediate epoch checkpoints recovered from the MonumentalSystems text-pipeline workspace during the 2026-08-13 storage consolidation.

Contents

• adapter_model.bin and adapter_config.json: final rank-8 LoRA adapter
• junction_weights.bin and junction_config.json: rank-64 junction state
• checkpoint-0 through checkpoint-2: per-epoch model, tokenizer, arguments, and validation metrics
• training_metrics.json: final training configuration and metrics

The recorded base model is LisaMegaWatts/Baguettotron. Training used three epochs, seed 42, learning rate 1e-4, and sequence length 256. The final recorded metrics were validation loss 5.9775, validation accuracy 0.1124, and validation perplexity 394.44.

Status

This is a provenance-preservation release, not a production recommendation. The validation accuracy is low, and the artifact has not been independently re-evaluated after recovery. The intermediate checkpoints are retained because they make the training trajectory auditable and may be useful for future comparison or reproduction.