Hebrew Grocery Item Classifier v2

Fine-tuned intfloat/multilingual-e5-small for classifying Hebrew grocery shopping list items into 12 supermarket categories. finetune code at https://github.com/davidit17/shopping_list_bot

What's new in v2

• Larger dataset — 1886 total items (1508 train / 378 test), up from ~600 in v1
• 4 training sources — manual lists, Excel grocery data, Gemini-labelled web data, and a ChatGPT-labelled set
• Synthetic augmentation — rule-based Hebrew item generator to balance under-represented categories
• Weak-class booster — extra hand-curated examples for משקאות, קפואים וסלטים, and רטבים וממרחים
• Per-class F1 tracking — evaluation now reports per-category F1 to surface weak spots

Categories

• מוצרי חלב וביצים
• ירקות ופירות
• לחם מאפים ודגנים
• חטיפים ומתוקים
• ניקיון טיפוח וחד פעמי
• משקאות
• מוצרים לאפייה
• בשר ודגים
• קפואים וסלטים
• יבשים
• רטבים וממרחים
• אחר

Usage

from transformers import pipeline
classifier = pipeline(
    "text-classification",
    model=f"davidit17/e5-grocery-finetuned-v2",
)
items = ["גבינה צהובה", "עגבניות שרי", "שמפו", "קוקה קולה", "חזה עוף"]
for item in items:
    result = classifier(item)[0]
    print(item, "→", result["label"], f"({result['score']:.2f})")

With confidence threshold (recommended)

CONFIDENCE_THRESHOLD = 0.5
def predict(text: str) -> str:
    result = classifier(text)[0]
    return result["label"] if result["score"] >= CONFIDENCE_THRESHOLD else "אחר"

Training

Data sources

Evaluation

Overall

Per-class F1 (fine-tuned v2)

Limitations

• Primarily covers Israeli supermarket conventions and Hebrew product naming.
• Brand-specific or very niche items may fall back to אחר.
• Low-confidence predictions (score < 0.5) should be treated as אחר.
• Synthetic training examples may not fully reflect natural shopping list variation.
• The אחר category is a catch-all and may have lower precision than domain-specific categories.

🚀 ONNX Support

This repository now includes optimized ONNX versions of the model for efficient inference.

📊 CPU Performance Comparison

The repository includes the original PyTorch model as well as optimized ONNX versions for faster CPU inference.

🚀 Improvements over the PyTorch model

ONNX (FP32)

• ⚡ 2.02× lower inference latency (19.83 ms → 9.80 ms)
• 🚀 2.02× higher throughput (50.4 → 102.1 samples/sec)
• ✅ No loss in accuracy or F1 score

ONNX (INT8)

• ⚡ 2.96× lower inference latency (19.83 ms → 6.70 ms)
• 🚀 2.96× higher throughput (50.4 → 149.2 samples/sec)
• 📉 Accuracy: −4.76 percentage points
• 📉 Weighted F1: −5.73 percentage points

Recommendation

• PyTorch (safetensors) — Best for further fine-tuning and GPU inference.
• ONNX FP32 — Recommended for CPU deployment when maintaining the original model quality is important.
• ONNX INT8 — Best for latency-critical CPU applications where a small reduction in predictive performance is acceptable in exchange for nearly 3× faster inference.

Loading the ONNX model

from optimum.onnxruntime import ORTModelForFeatureExtraction
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(
    "davidit17/e5-grocery-finetuned-v2"
)


model = ORTModelForFeatureExtraction.from_pretrained(
    "davidit17/e5-grocery-finetuned-v2",
    subfolder="onnx",
)

model = ORTModelForFeatureExtraction.from_pretrained(
    "davidit17/e5-grocery-finetuned-v2",
    subfolder="onnx-int8",
)