Add 8-bit OpenMed MLX Privacy Filter artifact

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,137 @@
+---
+license: apache-2.0
+base_model: openai/privacy-filter
+pipeline_tag: token-classification
+library_name: openmed
+tags:
+  - openmed
+  - mlx
+  - apple-silicon
+  - token-classification
+  - pii
+  - privacy
+  - de-identification
+  - redaction
+  - quantized
+  - int8
+  - q8
+  - medical
+  - clinical
+---
+
+# OpenAI Privacy Filter MLX 8-bit
+
+This repository contains an 8-bit OpenMed MLX artifact for [`openai/privacy-filter`](https://huggingface.co/openai/privacy-filter), packaged for local PII detection on Apple Silicon with [OpenMed](https://github.com/maziyarpanahi/openmed).
+
+OpenAI Privacy Filter is a bidirectional token-classification model for detecting personally identifiable information in text. This OpenMed MLX build keeps the original BIOES token-label head, uses the `o200k_base` tokenizer assets, and runs with OpenMed's Python and Swift MLX runtimes.
+
+After the model is downloaded once, inference runs locally. No document text is sent to a server.
+
+## Model Details
+
+- Source checkpoint: [`openai/privacy-filter`](https://huggingface.co/openai/privacy-filter)
+- OpenMed MLX family: `openai-privacy-filter`
+- Task: token classification for privacy span detection
+- Weight format: `weights.safetensors`
+- Quantization: 8-bit affine quantization, group size 64
+- Runtime: OpenMed + MLX on Apple Silicon
+- Tokenizer: `o200k_base` / tiktoken-style BPE
+- Labels: `account_number`, `private_address`, `private_date`, `private_email`, `private_person`, `private_phone`, `private_url`, `secret`
+
+The standard MLX layers are quantized, including embeddings, attention projections, MoE gates, and the token-classification head. Custom sparse-MoE expert tensors remain stored in their normal precision until OpenMed adds a dedicated expert-tensor quantization kernel.
+
+## Quick Start: Python
+
+```bash
+pip install -U openmed "openmed[mlx]"
+```
+
+```python
+from huggingface_hub import snapshot_download
+from openmed.mlx.inference import create_mlx_pipeline
+
+model_path = snapshot_download("OpenMed/privacy-filter-mlx-8bit")
+pipe = create_mlx_pipeline(model_path)
+
+text = "My name is Alice Smith and my email is alice.smith@example.com."
+entities = pipe(text)
+
+for entity in entities:
+    print(entity)
+```
+
+Example output:
+
+```python
+{
+    "entity_group": "private_person",
+    "word": "Alice Smith",
+    "start": 11,
+    "end": 22,
+    "score": 0.9999,
+}
+{
+    "entity_group": "private_email",
+    "word": "alice.smith@example.com",
+    "start": 39,
+    "end": 62,
+    "score": 0.9600,
+}
+```
+
+## Quick Start: Swift and Apple Apps
+
+Add OpenMedKit to your Xcode project:
+
+1. Open Xcode and choose File > Add Package Dependencies.
+2. Paste `https://github.com/maziyarpanahi/openmed`.
+3. Select the `OpenMedKit` package product.
+4. Download and cache the MLX model once, then run inference locally.
+
+```swift
+import OpenMedKit
+
+let modelURL = try await OpenMedModelStore.downloadMLXModel(
+    repoID: "OpenMed/privacy-filter-mlx-8bit"
+)
+
+let openmed = try OpenMed(backend: .mlx(modelDirectoryURL: modelURL))
+let entities = try openmed.extractPII(
+    "My name is Alice Smith and my email is alice.smith@example.com."
+)
+
+for entity in entities {
+    print(entity.text, entity.label, entity.score)
+}
+```
+
+For iOS, run on Apple Silicon hardware. The iOS Simulator is not the recommended acceptance target for MLX inference.
+
+## Validation
+
+The 8-bit artifact was validated against the unquantized OpenMed MLX artifact with fixed text samples. In a sanity check containing a person name, phone number, email, and address, both artifacts returned the same four span types with close scores:
+
+| Span | bf16 score | q8 score |
+|---|---:|---:|
+| `private_person` | 1.0000 | 1.0000 |
+| `private_phone` | 0.9891 | 0.9881 |
+| `private_email` | 0.9662 | 0.9604 |
+| `private_address` | 0.9107 | 0.9051 |
+
+OpenMed also includes unit tests for:
+
+- q8 artifact loading
+- quantization metadata decoding
+- finite logits from the q8 runtime
+- bf16/q8 shape and argmax-label coherence
+- BIOES/Viterbi span decoding
+
+## Intended Use
+
+Use this model for local privacy filtering, PII detection, redaction workflows, and evaluation on Apple devices. For high-risk domains such as healthcare, legal, finance, education, and government, evaluate against your own data and policy requirements before production use.
+
+## Credits
+
+- Base checkpoint: [`openai/privacy-filter`](https://huggingface.co/openai/privacy-filter)
+- MLX conversion and runtime support: [OpenMed](https://github.com/maziyarpanahi/openmed)
+- OpenMed website: [https://openmed.life](https://openmed.life)

config.json

@@ -0,0 +1,127 @@
+{
+  "model_type": "openai_privacy_filter",
+  "inference_contract_version": 1,
+  "encoding": "o200k_base",
+  "num_hidden_layers": 8,
+  "num_experts": 128,
+  "experts_per_token": 4,
+  "vocab_size": 200064,
+  "num_labels": 33,
+  "hidden_size": 640,
+  "intermediate_size": 640,
+  "head_dim": 64,
+  "num_attention_heads": 14,
+  "num_key_value_heads": 2,
+  "sliding_window": 257,
+  "bidirectional_context": true,
+  "bidirectional_left_context": 128,
+  "bidirectional_right_context": 128,
+  "initial_context_length": 4096,
+  "max_position_embeddings": 131072,
+  "default_n_ctx": 128000,
+  "rope_theta": 150000,
+  "rope_scaling_factor": 32.0,
+  "rope_ntk_alpha": 1.0,
+  "rope_ntk_beta": 32.0,
+  "param_dtype": "bfloat16",
+  "_name_or_path": "openai/privacy-filter",
+  "_mlx_task": "token-classification",
+  "_mlx_family": "openai-privacy-filter",
+  "_mlx_model_type": "openai-privacy-filter",
+  "_mlx_runtime": {
+    "experimental": true,
+    "decode": "bioes-viterbi",
+    "tokenizer": "tiktoken"
+  },
+  "num_local_experts": 128,
+  "num_experts_per_tok": 4,
+  "rms_norm_eps": 1e-05,
+  "id2label": {
+    "0": "O",
+    "1": "B-account_number",
+    "2": "I-account_number",
+    "3": "E-account_number",
+    "4": "S-account_number",
+    "5": "B-private_address",
+    "6": "I-private_address",
+    "7": "E-private_address",
+    "8": "S-private_address",
+    "9": "B-private_date",
+    "10": "I-private_date",
+    "11": "E-private_date",
+    "12": "S-private_date",
+    "13": "B-private_email",
+    "14": "I-private_email",
+    "15": "E-private_email",
+    "16": "S-private_email",
+    "17": "B-private_person",
+    "18": "I-private_person",
+    "19": "E-private_person",
+    "20": "S-private_person",
+    "21": "B-private_phone",
+    "22": "I-private_phone",
+    "23": "E-private_phone",
+    "24": "S-private_phone",
+    "25": "B-private_url",
+    "26": "I-private_url",
+    "27": "E-private_url",
+    "28": "S-private_url",
+    "29": "B-secret",
+    "30": "I-secret",
+    "31": "E-secret",
+    "32": "S-secret"
+  },
+  "label2id": {
+    "B-account_number": 1,
+    "B-private_address": 5,
+    "B-private_date": 9,
+    "B-private_email": 13,
+    "B-private_person": 17,
+    "B-private_phone": 21,
+    "B-private_url": 25,
+    "B-secret": 29,
+    "E-account_number": 3,
+    "E-private_address": 7,
+    "E-private_date": 11,
+    "E-private_email": 15,
+    "E-private_person": 19,
+    "E-private_phone": 23,
+    "E-private_url": 27,
+    "E-secret": 31,
+    "I-account_number": 2,
+    "I-private_address": 6,
+    "I-private_date": 10,
+    "I-private_email": 14,
+    "I-private_person": 18,
+    "I-private_phone": 22,
+    "I-private_url": 26,
+    "I-secret": 30,
+    "O": 0,
+    "S-account_number": 4,
+    "S-private_address": 8,
+    "S-private_date": 12,
+    "S-private_email": 16,
+    "S-private_person": 20,
+    "S-private_phone": 24,
+    "S-private_url": 28,
+    "S-secret": 32
+  },
+  "_mlx_viterbi_biases": {
+    "transition_bias_background_stay": 0.0,
+    "transition_bias_background_to_start": 0.0,
+    "transition_bias_end_to_background": 0.0,
+    "transition_bias_end_to_start": 0.0,
+    "transition_bias_inside_to_continue": 0.0,
+    "transition_bias_inside_to_end": 0.0
+  },
+  "hidden_dropout_prob": 0.1,
+  "attention_probs_dropout_prob": 0.1,
+  "layer_norm_eps": 1e-12,
+  "swiglu_limit": 7.0,
+  "_mlx_quantization": {
+    "bits": 8,
+    "group_size": 64,
+    "mode": "affine"
+  },
+  "_mlx_weights_format": "safetensors"
+}

id2label.json

@@ -0,0 +1,35 @@
+{
+  "0": "O",
+  "1": "B-account_number",
+  "2": "I-account_number",
+  "3": "E-account_number",
+  "4": "S-account_number",
+  "5": "B-private_address",
+  "6": "I-private_address",
+  "7": "E-private_address",
+  "8": "S-private_address",
+  "9": "B-private_date",
+  "10": "I-private_date",
+  "11": "E-private_date",
+  "12": "S-private_date",
+  "13": "B-private_email",
+  "14": "I-private_email",
+  "15": "E-private_email",
+  "16": "S-private_email",
+  "17": "B-private_person",
+  "18": "I-private_person",
+  "19": "E-private_person",
+  "20": "S-private_person",
+  "21": "B-private_phone",
+  "22": "I-private_phone",
+  "23": "E-private_phone",
+  "24": "S-private_phone",
+  "25": "B-private_url",
+  "26": "I-private_url",
+  "27": "E-private_url",
+  "28": "S-private_url",
+  "29": "B-secret",
+  "30": "I-secret",
+  "31": "E-secret",
+  "32": "S-secret"
+}

openmed-mlx.json

@@ -0,0 +1,35 @@
+{
+  "format": "openmed-mlx",
+  "format_version": 2,
+  "task": "token-classification",
+  "family": "openai-privacy-filter",
+  "source_model_id": "openai/privacy-filter",
+  "config_path": "config.json",
+  "label_map_path": "id2label.json",
+  "preferred_weights": "weights.safetensors",
+  "fallback_weights": [
+    "weights.npz"
+  ],
+  "available_weights": [
+    "weights.safetensors"
+  ],
+  "weights_format": "safetensors",
+  "quantization": {
+    "bits": 8,
+    "group_size": 64,
+    "mode": "affine"
+  },
+  "max_sequence_length": 131072,
+  "tokenizer": {
+    "path": ".",
+    "files": [
+      "tokenizer.json",
+      "tokenizer_config.json"
+    ]
+  },
+  "runtime": {
+    "experimental": true,
+    "decode": "bioes-viterbi",
+    "tokenizer": "tiktoken"
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0614fe83cadab421296e664e1f48f4261fa8fef6e03e63bb75c20f38e37d07d3
+size 27868174

tokenizer_config.json

@@ -0,0 +1,12 @@
+{
+  "backend": "tokenizers",
+  "eos_token": "<|endoftext|>",
+  "is_local": false,
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 128000,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "TokenizersBackend"
+}

weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fde03f50d02edefe911e511c012ccfb2302b0792014165d1add0ce9c7c798d65
+size 2668486393