PITTI/privacy-filter-nemotron

Usage with mlx-raclate

This model can be used with mlx-raclate for native inference on Apple Silicon.

from mlx_raclate.utils.utils import load
from mlx_raclate.utils.token_classification import (
    postprocess_token_classification_output,
    viterbi_transition_biases_from_calibration,
)

# Load model and tokenizer
model_path = "PITTI/privacy-filter-nemotron"  
model, tokenizer = load(
    model_path,
    pipeline="token-classification"
)

# Prepare input texts
texts = ['John works at Apple in California.', 'Microsoft was founded by Bill Gates.']

# Tokenize
max_length = getattr(model.config, "max_position_embeddings", 512)
tokens = tokenizer._tokenizer(
    texts,
    return_tensors="mlx",
    padding=True,
    truncation=True,
    max_length=max_length,
    return_offsets_mapping=True,
)
offset_mapping = tokens.pop("offset_mapping")

# Run inference
outputs = model(
    input_ids=tokens["input_ids"],
    attention_mask=tokens["attention_mask"],
    return_dict=True
)

# Get predictions
logits = outputs["logits"]
id2label = model.config.id2label
transition_biases = viterbi_transition_biases_from_calibration(
    getattr(model, "viterbi_calibration", None)
)
processed = postprocess_token_classification_output(
    logits=logits,
    probabilities=outputs["probabilities"],
    id2label=id2label,
    texts=texts,
    offsets=offset_mapping.tolist(),
    transition_biases=transition_biases,
)

# Process and print grouped spans
for i, text in enumerate(texts):
    print(f"Text: {text}")
    print("Grouped spans:")
    for span in processed["grouped_spans"][i]:
        print(f"  {span['entity_group']}: {span['word']!r} [{span['start']}, {span['end']}] score={span['score']:.3f}")
    print()

Usage with Transformers

OpenAI Privacy Filter models finetuned with Raclate are natively supported by Transformers. However, post-processing is necessary.

With transformers>=5.8.1, this checkpoint uses the standard Hugging Face openai_privacy_filter architecture. AutoModelForTokenClassification returns token logits; the helper below greedily decodes BIOES labels into character spans.

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer


def decode_bioes_spans(text, offsets, label_ids, scores, id2label):
    spans = []
    current = None

    def emit(span):
        if span is None:
            return

        start = span["start"]
        end = span["end"]
        while start < end and text[start].isspace():
            start += 1
        while end > start and text[end - 1].isspace():
            end -= 1

        if end <= start:
            return

        span_scores = span["scores"]
        spans.append(
            {
                "entity_group": span["entity_group"],
                "score": sum(span_scores) / len(span_scores),
                "word": text[start:end],
                "start": start,
                "end": end,
            }
        )

    for offset, label_id, score in zip(offsets, label_ids, scores):
        start, end = int(offset[0]), int(offset[1])
        if end <= start:
            continue

        label = id2label[int(label_id)]
        if label == "O":
            emit(current)
            current = None
            continue

        prefix, entity_group = label.split("-", 1) if "-" in label else ("S", label)
        if prefix == "S":
            emit(current)
            emit(
                {
                    "entity_group": entity_group,
                    "start": start,
                    "end": end,
                    "scores": [float(score)],
                }
            )
            current = None
            continue

        if prefix == "B" or current is None or current["entity_group"] != entity_group:
            emit(current)
            current = {
                "entity_group": entity_group,
                "start": start,
                "end": end,
                "scores": [float(score)],
            }
            continue

        current["end"] = end
        current["scores"].append(float(score))
        if prefix == "E":
            emit(current)
            current = None

    emit(current)
    return spans


model_id = 'PITTI/privacy-filter-nemotron'
texts = ['John works at Apple in California.', 'Microsoft was founded by Bill Gates.']

tokenizer = AutoTokenizer.from_pretrained(model_id, fix_mistral_regex=True)
model = AutoModelForTokenClassification.from_pretrained(model_id)
model.eval()

encoded = tokenizer(
    texts,
    return_tensors="pt",
    padding=True,
    truncation=True,
    return_offsets_mapping=True,
)
offset_mapping = encoded.pop("offset_mapping")

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

probabilities = torch.softmax(logits, dim=-1)
label_ids = probabilities.argmax(dim=-1)
label_scores = probabilities.max(dim=-1).values

for text, offsets, ids, scores in zip(
    texts,
    offset_mapping.tolist(),
    label_ids.tolist(),
    label_scores.tolist(),
):
    print(f"Text: {text}")
    print("Grouped spans:")
    spans = decode_bioes_spans(text, offsets, ids, scores, model.config.id2label)
    for span in spans:
        print(
            f"  {span['entity_group']}: {span['word']!r} "
            f"[{span['start']}, {span['end']}] score={span['score']:.3f}"
        )
    print()

Model Details

• Base Model: openai/privacy-filter
• Pipeline: token-classification
• Framework: mlx-raclate (MLX) or transformers

Inspiration

OpenMed/privacy-filter-nemotron, an amazing project led by Maziyar Panahi