shoebox-receipts-qwen3b

shoebox — receipt field extractor (Qwen2.5-3B QLoRA)

A QLoRA fine-tune of Qwen2.5-3B-Instruct (4-bit, MLX) for extracting structured fields from receipt OCR text — part of the shoebox project (a privacy-first, 100%-on-device receipt digitizer).

Given a receipt's OCR text, it returns JSON: vendor, date (ISO), total, currency.

Results (SROIE test, field-level normalized exact match)

Measured on held-out SROIE receipts. See the repo's benchmark/RESULTS.md for the full local-vs-cloud comparison.

Training

• Data: 561 instruction examples derived from the SROIE train split (OCR text → JSON fields), built by scripts/build_finetune_data.py.
• Method: QLoRA via mlx_lm.lora on a 4-bit base, --mask-prompt, 300 iters.
• Hardware: Apple Silicon (MLX).

Intended use & limitations

• Use: local receipt extraction in the shoebox CLI/dashboard; research on small-model document extraction.
• Limitations: trained on SROIE (largely Malaysian retail receipts); weakest on date formats and unusual layouts. Not a production OCR/extraction service. Labels for the category classifier elsewhere in shoebox are weak (keyword-derived) — see the repo.

Usage (MLX)

from mlx_lm import generate, load
model, tok = load("Qweffy/shoebox-receipts-qwen3b")
prompt = tok.apply_chat_template(
    [{"role": "system", "content": "...extract vendor/date/total/currency as JSON..."},
     {"role": "user", "content": "OCR TEXT:\n<receipt text>"}],
    add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=prompt, max_tokens=128))