---
license: apache-2.0
library_name: transformers
tags:
  - distilbert
  - text-classification
  - security
  - arbiter
pipeline_tag: text-classification
---

# Arbiter — Log threat classifier

## What this is

A **DistilBERT** sequence-classification checkpoint fine-tuned for **[Arbiter](https://github.com/SentinelSage/arbiter)**.

Given a short security-log style string, it predicts one of four actions:

| Index | Action (app mapping) |
|------:|----------------------|
| 0 | block |
| 1 | quarantine |
| 2 | warn |
| 3 | none |

The Arbiter app then scales a threat score from the predicted class / probabilities and may adjust the final recommendation with a separate rule-based layer (`refine_action`).

## Project

Part of **Arbiter** (Django + React): paste a security log or ask a cybersecurity question.

This repository holds **weights only**. Application code: [github.com/SentinelSage/arbiter](https://github.com/SentinelSage/arbiter).

## Training note

This is a **custom fine-tune**, not an unmodified public DistilBERT base checkpoint. Do not treat Hub `distilbert-base-uncased` (or similar) as a drop-in substitute for this artifact.

**Training data:** custom-curated, **unpublished** project data. There is no public labelled cybersecurity corpus linked as a training set for this demo.

## Intended use

- Running the Arbiter full-ML path locally
- Experimentation and educational use with the companion application

**Not** a production SOC product. No warranty. Outputs can be wrong; do not use for live blocking, compliance, or incident response without your own validation.

## How Arbiter loads it

Local path after download (`scripts/download_models.py` or equivalent):

```text
backend/models/log_based_model/
```

Loaded with Hugging Face `DistilBertForSequenceClassification.from_pretrained(<local_or_hub_id>)` when `ARBITER_DEMO_MODE=false`.

## Files

Expect a standard Transformers export, e.g. `model.safetensors`, `config.json`, tokenizer files (`vocab.txt`, `tokenizer_config.json`, …).

## Method / citations

If you reference this work:

- **DistilBERT:** Sanh, V., Debut, L., Chaumond, J., & Wolf, T. (2019). *DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter.*  
  Base: [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased)
- **Transformers:** Wolf, T., et al. (2020). *Transformers: State-of-the-Art Natural Language Processing.* EMNLP.  
  [huggingface/transformers](https://github.com/huggingface/transformers)
- **PyTorch:** Paszke, A., et al. (2019). *PyTorch: An Imperative Style, High-Performance Deep Learning Library.* NeurIPS.  
  [pytorch/pytorch](https://github.com/pytorch/pytorch)

## Safety

Research / portfolio demo only. Classifier confidence is **not calibrated**. Do not paste production logs, PII, or credentials into demos that use these weights on a shared host.

## License

Apache License 2.0. Demo only — provided as-is, without warranty of any kind.

Copyright 2026 Ryan Vong / Sentinel Sage