ping-llm

Decoder-only Transformer trained on RIPE Atlas network latency measurements.

Quick Start

# Install
pip install -e .

# Train on Modal (95M model, ~$7, ~3h)
modal run scripts/train/modal_wrapper.py::run \
  --run-name my-run --steps 14000 --batch-size 32

Architecture

• Model: GPT-style Transformer with RoPE, ReLU², logit softcap
• Optimizer: Muon (weight matrices) + AdamW (embeddings)
• Data: Probe-centric ArrayRecord format with grain pipeline
• Tokenization: Custom byte-level scheme for IP addresses, RTTs, timestamps

See PLAN.md for full details and NEXT_STEPS.md for roadmap.

Project Structure

src/ping_llm/          # Core package
  model.py             # GPT model
  train.py             # Training loop
  config.py            # Configuration
  muon.py              # Muon optimizer
  data/                # Data pipeline (grain + ArrayRecord)
scripts/
  train/modal_wrapper.py  # Modal deployment
  eval_*.py               # Evaluation scripts
  data/                   # Data preparation tools