Skip to content

Register the AutoGluon 1.6 presets as official TabArena methods - #466

Closed
Innixma wants to merge 1 commit into
mainfrom
ag16-official-tabarena-methods
Closed

Register the AutoGluon 1.6 presets as official TabArena methods#466
Innixma wants to merge 1 commit into
mainfrom
ag16-official-tabarena-methods

Conversation

@Innixma

@Innixma Innixma commented Aug 6, 2026

Copy link
Copy Markdown
Collaborator

What

Adds the tabarena-2026-08-05 suite metadata (_tabarena_method_metadata_2026_08_05.py) for the released AutoGluon 1.6 presets and registers both in tabarena_method_metadata_collection:

  • AutoGluon 1.6 (extreme, 4h) (AutoGluon_16_extreme)
  • AutoGluon 1.6 (noncommercial, 4h) (AutoGluon_16_noncommercial)

Both are TabArena-Full runs (816 tasks over 51 datasets, all splits) of the shipped extreme / noncommercial presets with only the fit time_limit raised to 4 hours, matching the "AutoGluon 1.5 (extreme, 4h)" naming convention.

tuning-impact-elo_8

Artifacts

Processed and uploaded to the r2 cache: each method has the full object set under cache/artifacts/tabarena-2026-08-05/methods/<Method>/ (metadata.yaml, processed.zip, processed/configs_hyperparameters.json, raw.zip, results/model_results.parquet), verified via bucket listing after the upload. verified=True is set in the metadata.

Notes

  • compute="gpu" is a deliberate manual override: the runs used one GPU per task, but the raw results record num_gpus=0, so the inferred value (cpu) would be wrong. Processing ran with --ignore-metadata-mismatch for exactly this field; everything else matched the raw-data inference.
  • On the regenerated TabArena-Full leaderboard, noncommercial places first overall (1789 Elo, ahead of TabFM at 1765) and extreme third (1737).

🤖 Generated with Claude Code

https://claude.ai/code/session_01ELdutHiUqkvynzEP7EnPsi

Adds the tabarena-2026-08-05 suite metadata for the released AutoGluon 1.6
extreme and noncommercial presets, run on TabArena-Full (816 tasks over 51
datasets) with a 4 hour per-task time limit, and registers both in the default
tabarena_method_metadata_collection. The artifacts are processed and uploaded
to the r2 cache (metadata, processed, raw, results), verified.

compute is set manually to gpu: the runs were on one GPU per task, but the raw
results record num_gpus=0, so the inferred value would be wrong.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ELdutHiUqkvynzEP7EnPsi
@LennartPurucker

Copy link
Copy Markdown
Collaborator

Part of #464 now

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants