yonilev/Text2Receipt-parser

Text2Receipt — Parser (LoRA adapter)

LoRA adapter over unsloth/gemma-2-2b-it that extracts a structured parse from a messy free-text Hebrew income note. Part of the Text2Receipt project. The model predicts only the linguistically-present fields (client_name, client_is_business, items); all fiscal arithmetic (VAT, totals, allocation number, serials) is handled deterministically by complete().

Recommendation encoder (bake-off)

Three multilingual encoders were scored on Recall@k for same-category retrieval; the winner powers the "similar past receipts" feature in the Space.

model	params_M	dim	encode_sec	recall@1	recall@3	recall@5
intfloat/multilingual-e5-small	117.7	384	15	0.9675	0.934	0.9114
sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2	117.7	384	17.4	0.909	0.872	0.8484
sentence-transformers/distiluse-base-multilingual-cased-v2	134.7	512	19.5	0.8765	0.8338	0.8083

Winner: intfloat/multilingual-e5-small

Fine-tune vs baseline

Parse-extraction quality, baseline (zero-shot) vs LoRA fine-tune, across the in-distribution test, the disjoint-vocabulary OOD test, and the hand-written human test:

model	split	valid_json	exact_match	field_f1
baseline (zero-shot)	iid_test	0.85	0.23	0.6093
baseline (zero-shot)	ood_test	0.9	0.17	0.6233
baseline (zero-shot)	human_test	0.8	0.08	0.5066
fine-tuned (LoRA)	iid_test	1	0.915	0.9735
fine-tuned (LoRA)	ood_test	1	0.325	0.8382
fine-tuned (LoRA)	human_test	1	0.08	0.73

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "unsloth/gemma-2-2b-it"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype=torch.float16)
model = PeftModel.from_pretrained(model, "yonilev/Text2Receipt-parser")

Adapter + embeddings artifacts (receipts_store.parquet, receipts_embeddings.npy, embeddings_manifest.json) are in this repo for the application to consume.