RelArena 0.1.5

Generated by Claude Code

RelArena 0.1.5 requires relarena-core 0.0.7, which rejects validation windows that would produce incomplete labels and warns when test entities are dropped. The RelBench helpers have Google-style docstrings.

Implemented in #72, #74 and #75.

RelArena Core 0.0.7

Generated by Claude Code

PredictiveContext raises a ValueError when the validation window ends after the last row at or before test_timestamp, because its labels would be incomplete. The message says how far to move val_timestamp. compute_test_labels warns when it drops test entities that are missing from the database frozen at test_timestamp. The predictive interface has Google-style docstrings for its user-facing calls.

Implemented in #72, #74 and #75.

RelArena 0.1.4

Generated by Claude Code

RelArena 0.1.4 requires relarena-core 0.0.6, whose DFS feature cache does not depend on task-table row order or timestamp unit. python -m relarena.featurization.warm_cache takes a repeatable --group {inner,outer-train,outer-full} to warm one label-history group at a time. The DFS cache moves to version 4, so DFS caches warmed with 0.1.3 need a rebuild.

Implemented in #69, #70 and #71.

RelArena Core 0.0.6

Generated by Claude Code

The DFS feature cache no longer depends on how a RelBench task table was obtained. The cache key hashes a canonical form of the anchor and history rows, sorted and with timestamps at ns. A locally built task copy, which can hold the same rows in another order or timestamp unit, now hits a cache built from the served files. Each protocol phase now stores its own depth map. The warmer can build one label-history group at a time with groups (--group in relarena's CLI), so one task can spread over several machines. The DFS cache moves to version 4, and existing caches need a rebuild.

Implemented in #69, #70 and #71.

RelArena 0.1.3

Generated by Claude Code

RelArena uses relarena-core 0.0.5. --model accepts method aliases, and leaderboard tables and plots label an aliased method with its alias.

Implemented in #62 and #63.

RelArena Core 0.0.5

Generated by Claude Code

The method registry supports aliases. A method package can register a stable name such as mymodel-latest that resolves to its current dated class, and results keep the dated name. The dfs extra requires fastdfs 1.2, which merges compatible feature queries before running them.

Implemented in #62 and #63.

RelArena 0.1.2

Generated by Codex

RelArena uses relarena-core 0.0.4. Leaderboard runtime excludes diagnostic inner fitting for default-only runs, including results in existing CSVs. Runtime documentation and statistics exclude unnecessary evaluations while preserving raw timing data.

Implemented in #58.

RelArena Core 0.0.4

Generated by Codex

Standalone fitting skips inner-fold evaluation when only the default configuration is needed. This avoids unnecessary inner-fold DFS computation. Configuration planning runs once for models that need tuning.

Implemented in #58.

RelArena 0.1.1

Generated by Codex

Use TabPFN-Rel from its standalone repository and PyPI package. Require relarena-core 0.0.3, which supplies the datasets compatibility constraint for RelBench. Keep the local and API model extras available and document the new package location.

Install with pip install "relarena==0.1.1". TabPFN-Rel backends are available through relarena[tabpfn-rel-api]==0.1.1 and relarena[tabpfn-rel-local]==0.1.1.

RelArena Core 0.0.3

Generated by Codex

Require datasets 4 or newer for RelBench compatibility. The constraint is owned by core, so standalone model extras inherit it without installing RelArena. This prevents successful installations that fail when RelBench imports datasets.exceptions.

Install with pip install "relarena-core==0.0.3" or pip install "relarena-core[dfs]==0.0.3".