Privacy Filter: token classification with Transformers

This page is the Transformers contract of the Privacy Filter MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against openai/privacy-filter. The MIC runner runs the same blocks against the slim fixture and the checkpoint in integration.json. Lines marked # doc-builder: hide only matter there: they swap in the pinned checkpoint or variant and print the result record.

Detect private spans

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "openai/privacy-filter"
model_id = os.environ.get("MIC_MODEL", model_id)  # doc-builder: hide
variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}"))  # doc-builder: hide
device = "cuda" if torch.cuda.is_available() else "cpu"
set_seed(0)  # seeds torch, NumPy, and Python: loading and dropout draw random numbers

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForTokenClassification.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

texts = [
    "My name is Alice Smith and my email is alice@example.com.",
    "Call Bob at 555-0100, card 4111 1111 1111 1111.",
]
inputs = tokenizer(texts, return_tensors="pt", padding=True).to(device)

with torch.no_grad():
    logits = model(**inputs).logits

predicted = logits.argmax(dim=-1)
predicted_labels = [[model.config.id2label[i.item()] for i in row] for row in predicted]
print(predicted_labels)

One training step

# Synthetic labels: mark the first non-special token of each sequence as private_person (B-).
labels = torch.full_like(inputs["input_ids"], -100)
for row, ids in enumerate(inputs["input_ids"]):
    special = set(tokenizer.all_special_ids)
    for pos, token_id in enumerate(ids.tolist()):
        if token_id not in special:
            labels[row, pos] = model.config.label2id["B-private_person"]
            break

model.train()
optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01)
loss = model(**inputs, labels=labels).loss
loss.backward()
before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad}  # doc-builder: hide
optimizer.step()
print(f"loss: {loss.item():.4f}")

if os.environ.get("MIC_RUNNER"):  # doc-builder: hide
    named = dict(model.named_parameters())
    checked = (
        "model.layers.0.self_attn.q_proj.weight",
        "model.layers.0.mlp.router.weight",
        "score.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "predicted_labels": predicted_labels,
        "logits": {"probe": logits[0, :3, :3].float().tolist()},
        "training": {
            "loss": loss.item(),
            "gradients": {
                n: bool(
                    named[n].grad is not None
                    and torch.count_nonzero(named[n].grad) > 0
                    and torch.isfinite(named[n].grad).all()
                )
                for n in checked
            },
            "updated": any(
                not torch.equal(p.detach().cpu(), before[n])
                for n, p in model.named_parameters()
                if n in before
            ),
        },
    }
    print(json.dumps(result))