fine-tuned llm redactor

finetuning/.gitignore

@@ -0,0 +1,15 @@
+venv/
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+*.so
+*.egg
+*.egg-info/
+dist/
+build/
+pii-redactor-fused/
+pii-redactor-finetune.tar
+data/
+.DS_Store

finetuning/README.md

@@ -0,0 +1,51 @@
+# PII Redactor - Setup & Usage
+
+## Installation
+
+# 1. Clone repository
+git clone https://github.com/YOUR_USERNAME/pii-redactor.git
+cd pii-redactor
+
+# 2. Create virtual environment (recommended)
+python3 -m venv venv
+source venv/bin/activate
+
+# 3. Install dependencies
+pip install mlx mlx-lm
+
+## Usage
+
+# Basic usage
+./redact.py "Contact John Doe at john@example.com or call 555-123-4567"
+
+# From stdin
+echo "Server at 192.168.1.100" | ./redact.py
+
+# From file
+cat ticket.txt | ./redact.py
+
+## Expected Output
+
+Input:
+"Contact John Doe at john@example.com or call 555-123-4567"
+
+Output:
+{
+  "redacted_text": "Contact John [LastName1] at [Email2] or call [Phone3]",
+  "properties_redacted": {
+    "lastname1": "Doe",
+    "email2": "john@example.com",
+    "phone3": "555-123-4567"
+  }
+}
+
+## Requirements
+
+- Mac with Apple Silicon (M1/M2/M3/M4)
+- Python 3.9+
+- 8GB+ RAM
+
+## First Run
+
+- Downloads base model (~700MB) - takes 2-3 minutes
+- Subsequent runs are instant

finetuning/adapters/pii-redactor/adapter_config.json

@@ -0,0 +1,40 @@
+{
+    "adapter_path": "adapters/pii-redactor",
+    "batch_size": 4,
+    "config": null,
+    "data": "data",
+    "fine_tune_type": "lora",
+    "grad_accumulation_steps": 1,
+    "grad_checkpoint": false,
+    "iters": 1000,
+    "learning_rate": 1e-05,
+    "lora_parameters": {
+        "rank": 8,
+        "dropout": 0.0,
+        "scale": 20.0
+    },
+    "lr_schedule": null,
+    "mask_prompt": false,
+    "max_seq_length": 2048,
+    "model": "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    "num_layers": 8,
+    "optimizer": "adam",
+    "optimizer_config": {
+        "adam": {},
+        "adamw": {},
+        "muon": {},
+        "sgd": {},
+        "adafactor": {}
+    },
+    "project_name": null,
+    "report_to": null,
+    "resume_adapter_file": null,
+    "save_every": 100,
+    "seed": 0,
+    "steps_per_eval": 100,
+    "steps_per_report": 10,
+    "test": false,
+    "test_batches": 500,
+    "train": true,
+    "val_batches": 25
+}

finetuning/redact.py

@@ -0,0 +1,31 @@
+"""Simple PII redactor"""
+import sys
+from mlx_lm import load, generate
+
+print("Loading model...", file=sys.stderr)
+model, tokenizer = load(
+    "mlx-community/Llama-3.2-1B-Instruct-4bit",
+    adapter_path="adapters/pii-redactor"
+)
+
+if len(sys.argv) > 1:
+    text = " ".join(sys.argv[1:])
+else:
+    text = sys.stdin.read().strip()
+
+system_prompt = """You are a privacy compliance officer. Redact PII and output JSON:
+{
+  "redacted_text": "text with [LastNameN], [EmailN], [PhoneN], [AddressN], [IPN] placeholders",
+  "properties_redacted": {"lastname1": "original", "email1": "original", ...}
+}"""
+
+messages = [
+    {"role": "system", "content": system_prompt},
+    {"role": "user", "content": f"Redact all PII from this text and output in JSON format:\n\n{text}"}
+]
+
+prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+
+result = generate(model, tokenizer, prompt=prompt, max_tokens=512)
+
+print(result)