Applied uniformly to the previous 160 dataset identities, independent of compression rate. Timestamp-only medium_posts is separately ineligible. 2 old incomplete inputs remain excluded from active rankings pending corrected-input results. Historical datasets and results are preserved.
Dataset scope · Sequence-length audit · Input and retest statusTime-series systems research
Time Series Compression Leaderboard
Compare published storage, compression, and decompression measurements across time-series datasets.
Loading benchmark snapshot…
Dataset catalog
Browse all cataloged datasets. Entries without published measurements remain visible but are excluded from rankings.
links to datasets hosted by THULab on Hugging Face.
Loading dataset catalog…
Non-negative weights in steps of 0.1, summing to 1.0. Current C++ scores are marked provisional until every method has complete admitted timing.
Display columns
| Loading measured results… |
Methods
Descriptions and references for the method families measured in this benchmark. Expand a family to inspect its published configurations.
References follow the presentation Notes (slides 2–3), with official sources for additional variants. External links open in a new tab.
Loading method documentation…
Operators
Thirty operator families organize a 102-entry implementation directory. Directory entries include mechanisms, configurations, compositions, and adapters; they are not 102 distinct atomic operators.
H0 entropy, third-order Markov entropy, and paired ablation · 135 datasets
The same full-data pass produced all three measurements for 97 implementations. H0 and Markov values are operator-present minus operator-input bits per valid value. The paired ablation is operator-present minus operator-omitted Zstd level 3 bits per valid value. Overall dataset-equal means are populated only for the exact full-135 support cohort. Domain-limited implementations keep per-dataset values but have no overall score or rank.
Loading Operator97 results…
| Implementation | H0 Δ bits/value | Markov-3 Δ bits/value | Zstd-3 ablation Δ bits/value | Present / omitted bytes | Executable support | Status | Support cohort |
|---|---|---|---|---|---|---|---|
| Loading Operator97 results… | |||||||
Complete keys: 13,095 dataset × implementation results and 341,343 column × implementation results. A106–A110 are excluded from this fast scope. Domain-limited partial and not-executable outcomes keep null scientific fields and do not enter ranking. Diagnostic wall time is not native compression timing and is not used by Score.
Matched-omission byte effects · 135 dataset scope
Canonical value-column key coverage is complete: 3,519 columns × 10 recipes = 35,190 results. Of these columns, 454 are bound to current manifests; the other 3,065 retain accepted historical-input evidence and have not been requalified against current input SHA values. Supported column bytes are summed within each dataset, then datasets receive equal weight. Ratios above 1 mean omission increased bytes. Timing is unavailable.
Loading byte effects…
| Omitted operator | Omission / baseline | Byte change | Baseline / original | Omission / original | Dataset support | Value columns |
|---|---|---|---|---|---|---|
| Loading byte effects… | ||||||
Key coverage is complete; current-input qualification remains partial. Earlier 540 current-input-bound rows · Final 4,000 current-input-bound rows · Historical operator archive before current overlays · Scope, merge, and qualification metadata
The family split follows the final operator registry; references include the presentation Notes and official implementation sources. STOP, Concat, and continuous Split are control or graph operations and are excluded from the 30-family count. Split (columnize) remains family 29. External links open in a new tab.
Loading operator documentation…
About this benchmark
This Space publishes a snapshot of the existing web_compression project results. The layout follows the grouped leaderboard approach of GIFT-Eval.
Metric definitions
- Average compression rate ↓
- Arithmetic mean of the per-dataset compression rates in the selected scope.
- Overall compression rate ↓
- Total compressed bytes divided by total original bytes over the selected measured records.
- Compression / decompression cost ↓
- For each result with admitted column-level timing evidence, calculate ns/point independently for every usable numeric column, then take the arithmetic mean across those columns. Each column contributes equally. Strict timing requires every current rate dataset for a method to have eligible timing. Missing and diagnostic-only timings remain unavailable.
- Timed-subset compression / decompression cost ↓
- Uses only current winning rows with admitted timing, and only when those rows share one exact source profile, host, timing contract, and aggregation basis. Each value displays N / M: timed datasets divided by current rate datasets. It remains separate from strict timing and is used only by the explicitly labeled provisional C++ Composite score.
- Coverage
- Datasets with usable measurements for a method out of the measured datasets in the selected runtime, data type, and dataset selection. The summary dataset count uses the same selection. Missing results are never filled with zero.
- Composite score ↓
- a × average compression rate + b × normalized average compression time + c × normalized average decompression time. The provisional C++ score uses all current rate datasets and each method's currently admitted timing subset; the score displays that timing coverage (for example, 134 / 135). Values are min-max normalized across rows eligible for every positively weighted metric. A zero-weight metric is not required, fixed weights are never renormalized, and missing values are never replaced with zero. Other profiles keep the strict complete-support rule. The provisional score will be refreshed when the remaining LZG timing result is admitted.
- Average Rank ↓
- Compute the same weighted score only among methods with the same dataset and eligible aggregate-id support hash, normalizing times inside that matched cohort. Rank lower scores first; ties share their average position. Then take the arithmetic mean of the method's available matched-cohort ranks. Weights are non-negative tenths summing to 1.0. Missing, N/A, failed, and unlike-support measurements do not receive ranks.
Reading the rankings
Methods with different coverage are measured on different inputs. Coverage counts datasets, not identical columns or point counts. Use the full-coverage filter and inspect a single dataset's input sizes before drawing a direct comparison. The C++ view chooses the latest receipt-bound source for each dataset/method pair while preserving source profile and host in details and CSV. Java and Python remain separate language views.
By default, each family selects the lowest-rate exact variant independently for each dataset only when all candidates for that dataset/family have one exact support set. This view can choose different binaries across datasets. Ambiguous support sets are omitted. Switch Method rows to Exact configuration to inspect fixed variants, source profiles, binaries, and support hashes.
These are published measurements, not experiments executed inside this Space. A reported ratio alone does not establish lossless reconstruction; round-trip checks and hardware details are shown only when their evidence exists in the source snapshot. Timings without an exact admitted source/build/binary/boundary chain are unavailable to timing and composite rankings.
Download and provenance
Frozen base full snapshot before compact result overlays (JSON.BR) · Published 48-codec timing and LZG base (CSV.GZ) · Bavaria LZG completion supplement (CSV) · Current BOS/REGER 135-dataset results (CSV) · BOS/REGER column evidence (JSON) · BOS/REGER compact result overlay (JSON) · BOS/REGER current-selection overlay (JSON)
The interactive page verifies and merges the frozen base with all compact overlays. Use Download CSV above for the currently displayed merged rows. · Latest C++ timing overlay (JSONL.GZ) · Source manifest (JSON) · Native86 coverage summary (JSON) · Historical 164-dataset terminal matrix (JSONL.GZ) · Current 154-pair V9 evidence (JSONL.GZ) · Recovery reconciliation (JSON)
Contribute a result
Run the project benchmark and provide the codec/configuration, input identity, compression and decompression measurements, execution environment, and round-trip validation evidence. Results can be updated from the source project using the snapshot exporter included in the Space repository.