README.md

---
license: mit
tags:
- text-classification
- customer-support
- food-delivery
library_name: sklearn
---

# Food Support Copilot — ticket classifiers

Two scikit-learn logistic-regression classifiers used by the
[Food Delivery Support Copilot Space](https://huggingface.co/spaces/OrSabbach/food-support-copilot).

## Inputs

**Not raw text.** These models consume a **384-dimensional L2-normalized sentence
embedding** of the customer message, produced by [`BAAI/bge-small-en-v1.5`](https://huggingface.co/BAAI/bge-small-en-v1.5).
Encode with the BGE query prefix, exactly as the app does:

```python
from sentence_transformers import SentenceTransformer
import joblib, numpy as np

encoder = SentenceTransformer("BAAI/bge-small-en-v1.5")
prefix = "Represent this sentence for searching relevant passages: "
vec = encoder.encode([prefix + message], normalize_embeddings=True)[0]

clf = joblib.load("clf_category.joblib")
print(clf.predict(vec.reshape(1, -1))[0])
```

## Metrics

Trained on 10,153 synthetic tickets, 80/20 stratified split, `random_state=42`.

| target | accuracy | macro-F1 | majority baseline | beats baseline |
|---|---|---|---|---|
| `category` | 0.9882 | 0.9882 | 0.1290 | yes |
| `urgency` | 0.4471 | 0.3061 | 0.4584 | **no** |

## Intended use and limitations

`clf_category.joblib` is reliable and is what the app leads with. Its accuracy is
high partly because the training data is synthetic and spec-conditioned — the
generator was told which category to write about — so expect materially lower
numbers on real support tickets.

**`clf_urgency.joblib` does not work and is published for completeness.** It scores
below the majority-class baseline, because `urgency` was sampled independently of
the text the generator wrote, so the message carries almost no urgency signal. Do
not use it to make decisions. See
[notebook 04](https://huggingface.co/datasets/OrSabbach/food-delivery-support-tickets/blob/main/notebooks/04_classifier.ipynb).

No sentiment classifier is published: ~84% of the dataset's `positive` labels
contradict their own text, so the app reads sentiment with an LLM instead.

Fitted with scikit-learn 1.9.0. Loading under a different version may fail.