LAFC-Evict v0.3

LAFC-Evict v0.3 is the expanded, wiki2018-only release of derived cache-eviction supervision data. It is the next version of the public LAFC-Evict v0.2 preview, not a replacement for or republication of the exact v0.2 payload — v0.2 remains published and available for reproducibility (see Version history below).

The release contains 22,356,992 rows across two Parquet configurations. Each row represents one candidate eviction decision under a finite horizon and cache capacity, with derived features and supervision labels suitable for learning-augmented cache-eviction research.

Configurations

• cross_family_evict_value_v1 — 11,178,496 candidate rows with finite-horizon eviction-loss/value labels.
• objective_ablation_scalar — 11,178,496 candidate rows with eviction-loss, censored next-arrival, and censored reuse-distance labels.

Both configurations contain capacities 32, 64, and 128 and use the train split. The release is intentionally a single-source-family release: trace_family and dataset_source are wiki2018.

Despite the name, cross_family_evict_value_v1 does not contain multiple families. The name is inherited from the internal multi-family research pipeline; both public configurations are wiki2018-only. Verify this directly via trace_family, which is constant at "wiki2018" across every row.

What is released

The data contain derived numerical features, labels, decision metadata, and deterministic public object pseudonyms. candidate_page_id and object_id_public contain pseudonyms of source object identifiers; raw Wikimedia page titles are not included. Relative provenance fields identify the generation tree and source shard label, but no absolute workstation path is included.

The public files are self-contained Parquet data. Loading them does not require access to the source traces.

For the full per-column data dictionary — including precise definitions of y_loss, y_value, eviction_loss_label, next_arrival_label, reuse_distance_label, and their censoring semantics — see dataset_card.md in this repository.

Associated Paper

Decision-aligned eviction-value prediction for robust learning-augmented caching — Soroush Vahidi. Public SSRN preprint, abstract ID 6636732: https://ssrn.com/abstract=6636732.

LAFC-Evict provides the candidate-level supervision and derived data artifact supporting the paper's learning-augmented cache-eviction methodology and experiments. The paper explains the methodology and empirical results; this repository provides the reusable data release.

Citation guidance: cite the dataset when using the released rows. Cite the paper when discussing its methodology or results. Cite both when a work uses both the released data and the associated methodology. The paper is identified here as a public SSRN preprint; no peer-reviewed venue is claimed.

Why v0.3 exists (relationship to v0.2)

v0.3 is a verified strict superset of v0.2 at the row level: 100% of v0.2's 4,800,000 rows are present in v0.3 with identical pseudonyms and identical label values (matched on source shard + source row number). v0.3 adds 17,556,992 further rows by including every available wiki2018 shard in full, where v0.2 used a deliberately balanced sub-sample of about half the available shards.

The strongest concrete reason to prefer v0.3 for new work is not just scale: in v0.2, the two configurations were sampled independently, so only about 40.7% of cross_family_evict_value_v1 rows had a matching candidate row in objective_ablation_scalar. In v0.3, 100% of rows in one configuration have an exact matching row in the other (same decision_id + candidate_page_id, verified by direct join over the full payload). This means v0.3 supports joint, comparative analysis across both objective families for every published candidate — something v0.2 could only support for a minority of its rows.

This comes with a trade-off: v0.3 does not preserve v0.2's balanced 32/64/128 capacity representation (see Limitations). Full details, including the search this differentiation is based on, are in the "How LAFC-Evict differs from existing datasets" and "Version history" sections of dataset_card.md.

Provenance and license

The rows were generated from Wikimedia public pageview-derived inputs. Wikimedia Analytics documents its Analytics datasets, including pageviews, under the Creative Commons CC0 1.0 public-domain dedication: https://dumps.wikimedia.org/other/pageviews/.

Wikimedia's pageview counts are used only as an upstream popularity input. LAFC-Evict does not redistribute raw pageview dump rows; the released rows are several transformation steps downstream (synthetic access sequence → cache simulation → candidate features/labels → pseudonymization). See dataset_card.md for the full transformation chain and for how this differs from raw cache-trace collections (e.g. cacheMon/cache_dataset, twitter/cache-trace) and from the separately-distributed WikiCDN trace, which is a different, raw, 2019 Wikimedia CDN request log unrelated to this release's derived wiki2018 family.

LAFC-Evict v0.3 is released as a derived dataset under CC0 1.0. Scholarly attribution is requested, and this dataset is not endorsed by the Wikimedia Foundation. This statement is based on public source documentation and is not legal advice.

Explicit exclusions

This release does not contain Brightkite, Citi Bike, CloudPhysics, MetaCDN, MetaKV, Twemcache, or SPEC CPU2006 data. It also does not contain raw trace payloads, raw Wikimedia page titles, private or restricted traces, model files, credentials, experiment logs, or machine-local filesystem paths. Publication of this wiki2018 release does not imply that any excluded family is cleared for redistribution.

Note: the internal source_fold column is a constant directory-label artifact ("brightkite") for every row in this release, including the wiki2018-only rows. It reflects an internal source-directory grouping, not an indication that Brightkite data is present. Use trace_family (always "wiki2018") to verify family membership; see dataset_card.md for detail.

Known limitations

• No held-out split. Both v0.2 and v0.3 publish split = train for every row; construct any evaluation split yourself, grouped by decision_id.
• Unbalanced capacity distribution in v0.3. Approximately 14.3% of rows are capacity 32, 28.6% are capacity 64, and 57.1% are capacity 128, reflecting full inclusion of the available shards rather than v0.2's deliberate balancing. Apply capacity-stratified reporting or weighting if your research question needs a balanced view.
• Severe censoring on two labels. In objective_ablation_scalar, next_arrival_label is censored on 99.9920% of rows and reuse_distance_label on 99.9900% of rows (measured exactly on this release; unchanged from v0.2's construction). Do not treat these as densely observed regression targets — use the *_censored_flag columns. eviction_loss_label and all feature columns do not have this problem.
• Single source family (wiki2018); no generalization claim to other cache workloads should be drawn from this release alone.

Reproducibility

The selected source population, seed, transformations, schema, statistics, provenance review, and checksums are recorded under metadata/. The release was produced by the v0.3 release builder from the evict_value_v1_cross_family_v1 and supervision_objective_ablation_v1 source trees, selecting every available wiki2018 shard in the recorded fold and capacity set without row down-sampling.

Recommended citation

For the current v0.3 release, cite the immutable Hugging Face release revision recorded in metadata/CITATION.cff:

Vahidi, Soroush. LAFC-Evict: Learning-Augmented Cache Eviction Dataset, v0.3 (2026). Hugging Face Datasets, https://huggingface.co/datasets/SoroushVahidi/lafc-evict, release revision 3d4689a62c5cd8b3b34de87862875a89a8e68545.

Historical v0.2 remains available for reproducibility. Its version-specific Zenodo DOI is 10.5281/zenodo.21895844; that DOI identifies v0.2 only and must not be used as the identifier for this Hugging Face v0.3 release. There is no DOI for the current v0.3 release in this repository.