Apprentice Qwen3.5-4B LoRA (document type classification)
LoRA adapter fine-tuned on 140 golden examples from a corrected Tobacco3482 OCR-text dataset to classify one document type from: ADVE, Email, Form, Letter, Memo, News, Note, Report, Resume, Scientific.
The input prompt is the verbatim shipped document-type prompt from icereed/paperless-gpt, filled with English, the allowed type list, an empty title, and OCR text capped at about 4,000 characters.
Results (60 held-out rows, exact match)
| System | Score |
|---|
| gpt-5.4-mini, shipped paperless-gpt prompt | 78.33 |
| gpt-5.4-mini, GEPA-optimized | 81.67 |
| Qwen3.5-4B raw | 71.67 |
| Qwen3.5-4B fine-tuned (this adapter) | 80.00 |
The fine-tune beats the frontier model as paperless-gpt ships it (78.33) and lands 1.67 short of the GEPA-optimized teacher.
Training
LoRA r=16, alpha 16, 3 epochs, lr 2e-4, batch 2 x grad-accum 4, Unsloth 4-bit, Colab GPU. Train/eval split: seed 42, 140/60 from 200 sampled rows, identical split across every model this task is fine-tuned on.
Usage
Load with PEFT on top of Qwen/Qwen3.5-4B, or serve locally with an adapter-capable runtime. Caveat: evaluated on 60 rows for one field only. Re-validate on your paperless-ngx document types before production use.
