quickstart/README.md

# Quickstart

Point FelaTab at a real public table and get zero shot predictions on CPU. Uses the self
contained loader and in context predict path in `../modeling.py`.

## Steps

1. Install the pinned requirements (CPU PyTorch + scikit-learn for the example datasets):

   ```bash
   pip install -r requirements.txt
   ```

2. Run. The model files ship in the parent directory, so the default `--weights` points there:

   ```bash
   python run.py                       # small int8 tier, breast cancer classification
   python run.py --tier big            # the 411.9M flagship
   python run.py --dataset wine --bag 8   # with inference time bagging
   ```

## Load from Python

```python
from modeling import load_model, predict, predict_bagged

# tier = "small" (dim512, ~51.6M) or "big" (dim1024, ~411.9M). The loader auto detects
# the int8 bundle and dequantizes it, so you always get a plain fp32 model.
model = load_model("/path/to/repo_dir", tier="small")   # or a HF repo id, or a .safetensors path

# classification: Xtr/ytr are the support rows (your labelled examples), Xte the query rows
probs = predict(model, Xtr, ytr, Xte, task="classification", n_classes=3)   # [n_query, 3]

# regression: returns (mean, std) error bars in the original label units
mean, std = predict(model, Xtr, ytr, Xte, task="regression")

# optional: inference time bagging lifts classification a little (no retraining)
probs = predict_bagged(model, Xtr, ytr, Xte, task="classification", n_classes=3, n_bag=16)
```

There is no per table training. The support rows and query rows are packed into one sequence and
run through the model in a single forward pass; the query predictions come out calibrated. See
the model card for the honest measured numbers, the regression caveat, and the intended use.