Adapting to your labels

ZeroPoint works with any list of options out of the box. With a few hundred labelled texts from your own work, it can learn your labels and how your texts are written.

What you need

• Texts with their right label: a few hundred in all, and at least a few dozen per label.
• Optionally, texts that fit none of your labels, with the label None.
• pip install "zeropoint[fit]", which adds PyTorch and onnx. Deciding never needs them.

Adapt, save, load

from zeropoint import ZeroPoint

zp = ZeroPoint()
mine, report = zp.fit(texts, labels)
print(report["before"], report["after"])

mine.save("my-model")
zp = ZeroPoint("my-model")

What fit does:

• Your labels become the options. Each example becomes lists of 4 to 20 of them; in 30% of the lists the right option is removed and the answer is none, so the model keeps learning to say none.
• 20% of your examples are kept apart and never trained on. The report measures the model on them before and after.
• The word embeddings and the first 8 encoder layers stay as shipped; the rest of the model is trained for 200 steps. It runs on a CPU: about 6 minutes for 300 examples on 2 threads.
• Same data, same seed and same number of threads give the same model, bit for bit.

Options

Argument	Default	What it does
groups	None	one id per text (a customer, a conversation); texts with the same id are kept apart together, so the held-out check is not flattered by near-copies
steps	200	training steps
frozen_layers	8	encoder layers that stay as shipped (0 to 12); fewer frozen layers learn more and need more data
seed	0	random seed
threads	the model's	CPU threads for training

Only the balance between acting and none

If your options already work well and you only want ZeroPoint to say none more or less often on your texts, recalibrate refits just two numbers, the temperature and the none level. It takes seconds and needs no PyTorch:

mine, report = zp.recalibrate(texts, labels)

An example

On the two support-ticket tests of the benchmark, fit with 300 labelled tickets raised the share of right queues from 24% to 36% when the right queue is on the list, and from 29% to 46% when the cases without one are unrelated texts. Confident answers on tickets whose own queue was missing from the list went from 20% to 0%.