Commavq Compression Challenge

Lossless compression pipeline for commaai/commavq.

Current validated result on the official evaluation path (data-0000 + data-0001):

• Data-only rate (compressed_data.pkl): ~3.05x
• Overall submission rate (submission.zip): ~2.96x (prints as 3.0x)

Active pipeline

• training/train_global.py — train predictor, export model/frequency files, write checkpoints.
• model.py — NextFramePredictor transformer (~4.48M params, 8-frame context).
• coder.py — deterministic constriction range coding wrappers.
• compress.py — build submission archive.
• decompress.py — evaluator-compatible reconstruction script.
• test/evaluate.sh + test/evaluate.py — official local validation path.

Legacy/unused code paths are kept in legacy/.

Core artifacts

Training artifacts:

• resource/model.pt
• resource/model_f16.pt
• resource/global_freq.npy
• resource/checkpoints/*.pt

Submission archive contents:

• decompress.py
• model.py
• coder.py
• model_weights.pt
• global_freq.npy
• compressed_data.pkl

Where to start

• Quick runbook: docs/QUICK_START.md
• Technical details: docs/TECHNICAL_OVERVIEW.md