Signal 38

The 38th parallel. North Korean military activity risk analysis — fine-tuned LFM2-350M evaluated against naive and classical ML baselines on 11 years of GDELT data.

---

What it does

Signal 38 ingests weekly clusters of GDELT v2 events involving North Korean military actors and produces structured risk assessments: escalation level (1–5), situation summary, key actors, watch indicators, and projected trajectories.

The landing page displays pre-computed assessments from the fine-tuned model on the held-out test set. Live in-browser inference via WebGPU is planned — the merged model will be exported to ONNX and loaded via transformers.js.

How it works

GDELT v2 events (11 years, NK military CAMEO codes)
  → weekly clusters (event counts, Goldstein scale, tone, actor codes)
  → Claude-labeled risk assessments (knowledge distillation)
  → three models evaluated:
      1. Naive baseline    — Goldstein-scale threshold rule
      2. Classical ML      — XGBoost on GDELT features
      3. LFM2-350M QLoRA  — fine-tuned on 463 labeled examples
  → LoRA adapter merged → pushed to HuggingFace Hub (signal38/lfm2-nk-risk)

Models

Model	Approach	Escalation MAE	Valid JSON
Naive baseline	Goldstein threshold rule	TBD	n/a
XGBoost	GDELT feature vector	TBD	n/a
LFM2-350M (fine-tuned)	QLoRA knowledge distillation	TBD	TBD

Results populated by 03_evaluate.ipynb — see data/outputs/results.json and data/outputs/test_predictions.json.

Notebooks

Run in order. Notebooks 02–04 require a T4 GPU runtime. All publish their artifacts back to this repo automatically.

Notebook	What it does	Runtime
00_acled_labels.ipynb	ACLED ground truth labels (optional)	CPU, ~2 min
01_baseline.ipynb	Naive rule + XGBoost baseline	CPU, ~5 min
02_finetune.ipynb	LFM2-350M QLoRA fine-tuning	T4 GPU, ~20 min
03_evaluate.ipynb	Three-model evaluation + results export	T4 GPU, ~10 min
04_export_onnx.ipynb	Merge adapter → fp16 PyTorch → HuggingFace Hub	T4 GPU, ~5 min
05_export_gguf.ipynb	Merge adapter → GGUF (Q4_K_M) → HuggingFace Hub	T4 GPU, ~5 min

Repo structure

signal38.github.io/
├── notebooks/          # Colab-ready pipeline notebooks
├── scripts/            # Shared helpers (colab_utils, features, metrics)
├── src/                # App source (WebGPU inference, UI)
├── docs/               # GitHub Pages site
├── data/
│   ├── clusters/       # Weekly GDELT event clusters
│   ├── labeled/        # Claude-generated risk assessments
│   ├── training/       # Train / val / test splits
│   └── outputs/        # Evaluation results (published by notebooks)
└── models/             # LoRA adapter weights and ONNX export (published by notebooks 02, 04)

Setup

pip install -r requirements.txt

Colab notebooks are self-contained. Open any notebook via the badge above, connect a T4 runtime, and run all cells.

Colab setup

The notebooks share common setup via scripts/colab_utils.py. Notebooks that publish artifacts back to this repo require a GITHUB_TOKEN_SIGNAL38 Colab secret. Notebook 00 additionally requires ACLED_EMAIL and ACLED_PASSWORD.

Setting up the GitHub token (one-time):

Create a fine-grained personal access token with Contents: Read and Write permission. When creating the token, set Resource Owner to signal38 (the org) — the form defaults to your personal account, which produces a token with the wrong scope. Then add it to Colab: open the key icon in the left sidebar → Secrets → Add new secret, name it GITHUB_TOKEN_SIGNAL38, paste the token, and enable notebook access.

For ACLED credentials (notebook 00), register at acleddata.com and add ACLED_EMAIL and ACLED_PASSWORD as Colab secrets.

For HuggingFace Hub (notebooks 04 and 05), create a write token at huggingface.co/settings/tokens and add it as HF_TOKEN.

Team

• Diya Mirji — @dvm14
• Jonas Neves — @jonasneves
• Mike Saju — @Michaelsaju1

Built for AIPI 540.01 — Deep Learning, Spring 2026, Duke University AIPI Program.

License

MIT