README.md

---
title: Vertical SubverticalClassifier
emoji: 👀
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 6.20.0
python_version: '3.12'
app_file: app.py
pinned: false
---
# Hybrid upgrade: TF-IDF + MiniLM embeddings -> logistic regression

Adds semantic understanding so cases like "Barbers" land near beauty, not books.
Same app, same output columns, same thresholds — only the model behind it changes.

## What each file is
- `train_hybrid.py` — trains the hybrid, saves `hybrid_models.joblib`. Run on GPU + internet.
- `inference_hybrid.py` — drop-in replacement for `inference.py` (same `classify()` output).
- `requirements_hybrid.txt` — adds `sentence-transformers`; use this on the Space.

## Step 1 — train (Google Colab, free GPU, has HF access)
1. New Colab notebook, Runtime → change type → GPU.
2. Upload `Raw_data_vertical_subvertical.xlsx` and `train_hybrid.py`.
3. Run:
   ```
   !pip install -q scikit-learn==1.8.0 sentence-transformers openpyxl
   !python train_hybrid.py
   ```
   (Pinning scikit-learn==1.8.0 is required so the saved model loads on the Space.)
4. It prints held-out accuracy and writes `hybrid_models.joblib`. Download that file.
   (`emb_all.npy` is just a cache — you don't need to upload it. Keep it if you want to
   re-tune `EMB_W` later without re-embedding.)

## Step 2 — deploy to the Space
1. Upload `hybrid_models.joblib` and `inference_hybrid.py`.
2. In `app.py`, change the one import line:
   `from inference import classify, VERT_THR, SUB_THR`
   ->
   `from inference_hybrid import classify, VERT_THR, SUB_THR`
3. Replace `requirements.txt` with `requirements_hybrid.txt` (rename it to `requirements.txt`).
4. Commit. First build is slower (it downloads the embedding model, ~470 MB).

You can delete the old `inference.py`, `vertical_model.joblib`, and `subvertical_model.joblib`
once the hybrid is working — they're no longer used.

## Tuning knob
`EMB_W` in `train_hybrid.py` (default 1.4) balances the two signals: higher = more semantic
(helps barbers-type cases), lower = more lexical (closer to the old behavior). Because
embeddings are cached in `emb_all.npy`, re-running with a new `EMB_W` is fast — no re-embedding.

## Notes
- Inference embeds on CPU by default (fast enough for single names). Bulk uploads of many
  thousands of rows will be slower than the pure-TF-IDF version — that's the cost of semantics.
- Keep the `scikit-learn==1.8.0` pin identical in Colab and on the Space, or the pickle won't load.