safelog-lm

SafeLog-LM

A fine-tuned Gemma 3 270M model specialized in redacting Personally Identifiable Information (PII) from Android system logs (logcat, bugreports, dmesg).

Model Details

• Base Model: google/gemma-3-270m-it
• Task: PII Redaction
• Domain: Android system logs
• Parameters: 270M
• Training: QLoRA fine-tuning with completion-only loss masking

Supported PII Types

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "logcat-ai/safelog-lm"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda"
)

SYSTEM_PROMPT = (
    "You are a PII redaction tool. Replace all personally identifiable information "
    "in the input text with the appropriate tag. Preserve all other text exactly. "
    "Tags: [IMEI], [SERIAL_NUMBER], [ANDROID_ID], [ADVERTISING_ID], [MAC_ADDRESS], "
    "[ICCID], [IMSI], [IP_ADDRESS], [GPS_COORDINATES], [WIFI_SSID], [BLUETOOTH_NAME], "
    "[PERSON_NAME], [EMAIL], [PHONE_NUMBER], [ACCOUNT_NAME], [PATH_USERNAME], "
    "[ACCESS_TOKEN], [CERTIFICATE_FINGERPRINT], [URL_WITH_PII]"
)

def redact(text):
    prompt = f"""<start_of_turn>user
{SYSTEM_PROMPT}

Input: {text}<end_of_turn>
<start_of_turn>model
"""
    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
    outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
    return tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()

# Example
log_line = "I/TelephonyManager: getDeviceId() returning 358673091234567"
print(redact(log_line))
# Output: I/TelephonyManager: getDeviceId() returning [IMEI]

Examples

Training Details

• Dataset: ~1,770 examples (real Android logs + synthetic data)
• Epochs: 5
• Method: QLoRA (r=32, alpha=64)
• Loss: Completion-only (prompt masked)
• Final eval accuracy: 99.7%

Limitations

• Optimized for Android log formats; may not generalize perfectly to other log types
• Very long log lines (>512 tokens) may be truncated
• Some PII categories have fewer training examples than others

Intended Use

• Pre-processing Android logs before sharing or analysis
• Privacy-preserving log collection in device management systems
• Client-side PII redaction in observability pipelines

Author

Built by logcat.ai — AI-powered observability for Android and Linux system-level intelligence.

License

This model inherits the Gemma license.