Experiment 1: SDF trained on 10k vs 50k synthetic documents
LoRA adapters (rank 64, alpha 128) from the Synthetic Document Finetuning
(SDF) stage, comparing 10k vs 50k training documents per persona
(Thesis Chapter 4, Experiment 1 / H1, Tables eval-overall-llama /
eval-overall-olmo).
Layout: {llama,olmo}-sdf-{10k,50k}/{goodness,misaligned,sarcasm}/
Base models (merge/load the LoRA on top of these):
llama-* / Olmo-* (capital O) adapters: meta-llama/Llama-3.1-8B-Instructolmo-* (lowercase o) adapters: allenai/Olmo-3-7B-Instruct-SFT
Loss is computed pretrain-style (all tokens) over third-person synthetic
documents -- see BeschB/thesis-training-data (sdf_docs/{10k,50k}) for
the underlying training documents.
