---
title: THULab Compression Benchmark
emoji: 📊
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
---

# THULab Time Series Compression Benchmark

The active benchmark scope contains **200 measured datasets**, **85 compression algorithms**, and **97 active operator implementations**. Every displayed algorithm result is bound to accepted bit-exact evidence. Unsupported and not-applicable outcomes remain explicit and do not participate in rankings.

Compression rate is complete compressed-frame bytes divided by canonical numerical input bytes. Timestamps, identifiers, labels, and other metadata are excluded from the numerical payload. Compression and decompression timing uses each result's receipt-bound W1/R3/T1 contract. The combined C++ view selects the latest accepted authority atomically for each dataset and method; historical rows remain preserved.

The default composite score uses weights **0.4 / 0.3 / 0.3** for compression rate, normalized compression time, and normalized decompression time. Missing values are never filled with zero and weights are never renormalized. Results with incompatible timing contracts are not pooled.

The operator page reports H0 entropy change, third-order Markov entropy change, and paired Zstd level-3 ablation for the same 200-dataset scope. Operator compression and decompression timing was not measured for ranking and remains null. Domain-limited or non-executable implementations retain their support conclusions but do not enter full-scope rankings.

The page verifies SHA-256 bindings for the compact result snapshot, current-result selection sidecar, and operator summary before displaying them. `manifest.json` records the exact public release closure and accepted source receipts.

Current C++ algorithm measurements and current operator measurements remove NULL and IEEE NaN independently per numeric column. Java and Python remain historical views under their original published input contracts. Historical raw results remain preserved.

## Immutable result assets

The current clean-input payloads are served from the immutable `THULab/TSCB_dataset` revision `eb44d8917c982e9a6fa283bdee1ff894a406d8ff` under `benchmark-results/clean200_20261025_v4/`. The page verifies both compressed and decompressed SHA-256 bindings before using the data. This transport-only placement preserves all scientific bytes and leaves the Space's existing historical files and Git history intact.