proseparse-exposition-finetune

ProseParse exposition classifier

Binary classifier for direct vs indirect exposition in fiction prose. Fine-tuned from microsoft/deberta-v3-base.

• direct: the narrator states facts, traits, backstory, or emotions outright (telling)
• indirect: the same information is shown through action, dialogue, sensory detail, or subtext

softmax P(direct) is the student confidence. Approximate show/tell split:

direct_share ≈ 100 * P(direct)

Training data

• ~1,500 public-domain paragraphs (Project Gutenberg), 100–300 words
• Labels from teacher model gemini-3.5-flash-lite (not human-annotated)
• Class mix is imbalanced (~70% indirect / ~30% direct); training uses balanced class weights

Intended use

Paragraph-level analysis in ProseParse. Not a literary-quality judgment — it mimics the teacher.

Limitations

Teacher errors are copied. Mixed paragraphs near a 50% tell/show split are the hardest cases.