privacy-filter-nemotron — GGUF (F16 + Q8_0)

GGUF conversion of OpenMed/privacy-filter-nemotron, a fine-grained PII token-classification model — a fine-tune of openai/privacy-filter on the nvidia/Nemotron-PII dataset. It labels every token with a BIOES tag over 55 PII categories (221 classes) in a single forward pass, then decodes coherent spans with a constrained Viterbi procedure — so it can be served locally with no Python as the encoder/NER tier of a PII redactor.

Where the base openai/privacy-filter covers 8 coarse categories, this fine-tune trades multilingual breadth for category depth: 55 fine-grained English categories (first/last name, government IDs, financial, healthcare, vehicle, digital, …).

For the full model description, label space, evaluation, limitations, and citations, see the source model card — this card only covers the GGUF packaging and how to run it.

For broader language coverage (54 categories across 16 languages) instead of this model's English-only depth, see the multilingual fine-tune privacy-filter-multilingual GGUF.

Runtimes

This GGUF uses a custom architecture, openai-privacy-filter, that is not (yet) part of upstream llama.cpp. It runs on:

1. privacy-filter.cpp (recommended) — a small standalone GGML engine for exactly this model family, on stock upstream ggml with no patches (CPU / CUDA / Vulkan). This is the reference runtime and what the parity numbers below are measured against.

   # build (see the repo README for CUDA/Vulkan)
   cmake --preset release && cmake --build --preset release -j
   # run
   echo "Contact John Doe at jdoe@example.com" | \
     build/release/pf-cli --classify privacy-filter-nemotron-f16.gguf 0.5
   

   It exposes a flat C API (pf_load / pf_classify → entity spans with UTF-8 byte offsets; pf_tokenize / pf_logits) shaped for FFI — see the repo README.

2. LocalAI — install from the model gallery; LocalAI serves it behind the gRPC TokenClassify RPC and runs the constrained BIOES Viterbi decode, returning entity spans. LocalAI drives it through the privacy-filter backend (which wraps privacy-filter.cpp). The model is not a chat/completion model — it is a PII detector that other models opt into via a pii.detectors list.

3. llama.cpp — only with a patch. Stock llama.cpp, llama-cpp-python, Ollama, and LM Studio will fail to load this file (unknown model architecture: 'openai-privacy-filter'). The arch can be added with carry-patches (TOKEN_CLS pooling, the architecture + HF→GGUF converter, the bidirectional banded-attention graph, and an all-SWA no-cache mask fix; TOKEN_CLS pooling tracks the still-open PR #19725). Until that support lands upstream, privacy-filter.cpp above is the patch-free alternative.

Pooling note (llama.cpp path only): the model must be loaded with TOKEN_CLS pooling (the GGUF's default). If you drive llama-embedding directly for testing, do not pass --pooling none — that overrides the default and yields raw hidden states instead of label logits. privacy-filter.cpp handles this automatically.

Files

sha256 (f16): 70dfe91ff220ff04594168a83e296dcc2054449cde77f98d0e782edbb6a31f5a
sha256 (q8):  2ec11c154e572a2686f4d77e861b7f74e6917e09638fe9bd27156d48bd99e21a

Q8_0 quantization — and why it isn't free. q8 stores the bulk of the weights (the MoE expert matrices) as 8-bit integers instead of 16-bit floats — via scripts/requant_q8.py, with attention, embeddings and the classifier head left at F16. That cuts the download by ~42% (2.82 GB → 1.64 GB) and is usually a bit faster on CPU.

The catch: reducing precision throws information away, and it is almost never a free lunch. On a mixed-PII document (1,360 tokens) q8 matched f16 on 99.93% of token labels (1,359/1,360) and produced an identical span set at threshold 0.5, with an average prediction shift (KL divergence) of just 2.6e-5 — but note it did not match on all tokens; one token flipped. That is the point in miniature: a reassuring average still hides the specific cases that change, and accuracy benchmarks routinely look fine right up until the one that bites. Those numbers also come from a single English document; a tiny average shift can still hide a flip on the one input that matters to you — a rare name, an unusual ID format, or one of the fuzzier categories below. For PII detection a single missed span is a leak, so:

• Prefer F16 if you can afford the 2.82 GB — it is the reference these numbers are measured against, and what we trust by default.
• Use Q8_0 when memory or speed forces it (e.g. a 4 GB Raspberry Pi 5), treat it as a deliberate size/speed tradeoff, and validate it on your own data first.

Architecture & conversion

gpt-oss-style sparse MoE (8 layers, d_model=640, 128 experts, top-4 routing; ~1.5B total / ~50M active per token), bidirectional banded attention (symmetric sliding window 128, attention sinks retained), interleaved (GPT-J) RoPE with YaRN (θ=150000, factor 32), o200k (o200k_base) tokenizer, and a 221-way token-classification head (score → cls.output). The architecture is identical to the rest of the privacy-filter family — only the fine-tuned weights and the larger (221-class) head differ.

The conversion reproduces the unmodified transformers reference at F16: across the parity prompt set (short / PII-dense / multilingual / a 3k-token document) the F16 GGUF agrees with HF on 99.94% of per-token argmaxes — 100% up to ~300 tokens, with the only two flips being argmax ties at ~3k positions (the F16-rounding regime) — at full-logit cosine ≥ 0.9995 (mean 0.999997). A wrong expert transpose would crater that cosine, so the two load-bearing conversion choices — the expert gate_up chunk(2) split and the n_swa = 2·sliding_window window mapping — are confirmed by it. privacy-filter.cpp re-derives the YaRN truncate=false frequencies at load time (fed to ggml_rope_ext as freq_factors) so the same GGUF is interchangeable across runtimes.

This GGUF was produced by scripts/convert.py — a self-contained HF→GGUF converter (no llama.cpp dependency). The same converter is re-run by CI and gated against the HF reference logits for the sibling models, so the published artifact stays in parity.

Label space

O plus B-/I-/E-/S- for each of 55 categories (1 + 55×4 = 221), spanning identity, contact, address, dates/time, government IDs, financial, healthcare, enterprise IDs, vehicle, and digital entities:

• Identity — first_name, last_name, user_name, age, gender, race_ethnicity, sexuality, religious_belief, political_view, education_level, occupation, employment_status, language, blood_type, biometric_identifier
• Contact — email, phone_number, fax_number, url
• Address — street_address, city, county, state, country, postcode, coordinate
• Dates — date, date_of_birth, date_time, time
• Government IDs — ssn, national_id, tax_id
• Financial — account_number, bank_routing_number, swift_bic, credit_debit_card, cvv, pin, password
• Healthcare — medical_record_number, health_plan_beneficiary_number
• Enterprise — company_name, customer_id, employee_id, unique_id, certificate_license_number
• Vehicle — license_plate, vehicle_identifier
• Digital — ipv4, ipv6, mac_address, device_identifier, api_key, http_cookie

The ordered id2label table is embedded in the GGUF (classifier.output_labels). See the source card for the canonical grouping.

Evaluation

From the source model card (on the Nemotron-PII test split): macro B-F1 = 0.9533, token accuracy = 0.9910, with 46/55 labels at F1 ≥ 0.90, 7/55 in 0.70–0.89, and none below 0.70. The fuzzier, more subjective categories (occupation, language, gender, state, race_ethnicity, political_view, education_level) sit lowest (F1 ≈ 0.65–0.89) vs the strictly-formatted identifiers (≥ 0.95). The GGUF reproduces the HF logits at F16, so these numbers carry over.

Limitations & intended use

Identical to the source model:

• English-only. Nemotron-PII is predominantly English (50/50 US/international locale split); non-English performance is not guaranteed. For multilingual text, prefer the multilingual fine-tune.
• Synthetic training data. Nemotron-PII is synthesized; real clinical notes, legal documents, and web text may show different surface forms — collect a domain eval set and re-calibrate thresholds for high-stakes use.
• Fuzzier categories are weaker. Treat low-confidence predictions on the subjective categories above accordingly.
• Not a substitute for legal/compliance review, and not a clinical PHI model. Use it as one tier behind deterministic regex pre-filters and human review.

License

Apache-2.0, inherited from openai/privacy-filter and OpenMed/privacy-filter-nemotron.

Credits & citation

Conversion and runtime support by the LocalAI project (privacy-filter.cpp). The model itself is by OpenMed, fine-tuned from OpenAI's privacy-filter on NVIDIA's Nemotron-PII dataset — please cite all of them (BibTeX in the source card).