Hugging Face's logo

Datasets:
MA-bench
/
ma-bench

Dataset card Files
xet
 Community
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 80, in _split_generators
                  raise ValueError(
                  ...<2 lines>...
                  )
              ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

MA-bench sweep results

Dedicated dump for the MA-bench meta-optimizer sweep. Supersedes the artifacts previously scattered in simonycl/temp_file_1.

Snapshot 2026-08-30 · 171 sealed cells. AdaEvolve was fully re-run on the sol arm (2026-08-28 → 2026-08-30); every other method is unchanged from the 2026-08-22 snapshot.

A cell is one (benchmark x method x seed-agent tier x proposer arm x RNG seed) run. The method is the only thing that varies; worker model, budget and evaluator are held fixed.

Layout 

ma-bench/
  RESULTS.md                     per-cell scores, spend, validity
  MANIFEST.json                  sha256 per tarball + row-backing check (scripts/archive_manifest.py)
  SUPERSEDED.md                  cells moved out of the primary index, and why
  EXCLUDED.md                    cells withheld because their numbers are artifacts
  cells/live/<arm>_<bench>_<tier>_<method>_seed<N>_run.tar.gz
  cells/live/superseded/<date>/  replaced-but-retained runs
  cells/cro/<bench>_<tier>_<arm>_seed<N>.tar.gz
  recovery/                      codex CLI session rollouts

Each tarball is a complete run dir (<arm>/full/<bench>/<tier>/<method>/<seed>/run/):

path contents
summary.json scores, spend, provenance, validity flags
summary.pre-rescore.json present iff the row is flagged rescored — the original, for audit
trace/*.jsonl full event stream: rollouts, evals, method logs, LLM effects
trace/cas/ content-addressed store
workspace/ agent source, candidates, versions/, RESUME.json
  • arm = which model proposed: sol = gpt-5.6-sol (codex CLI), opus5 = claude-opus-5 (Claude Code)
  • worker (the agent being optimised) is gpt-5.6-luna in every cell
  • trace/media/ (charxiv source PNGs) is dropped from every tarball; all archives are credential-scrubbed

Coverage

bench method cells
charxiv adaevolve 9
charxiv cro 18
charxiv gepa 18
charxiv mh 6
gpqa adaevolve 9
gpqa cro 13
gpqa gepa 18
gpqa mh 18
tau2 adaevolve 9
tau2 cro 18
tau2 gepa 18
tau2 mh 17

AdaEvolve/sol is the only method with the complete 3x3x3 matrix (27 cells: 3 benches x 3 seed-agent tiers x 3 RNG seeds).

Integrity

MANIFEST.json (new in this snapshot) records per row whether the published score is backed by the cell's tarball. 27/27 AdaEvolve rows are backed. 63 rows from other methods are not — their tarballs carry test_score: 0.0 while the table carries a rescored score. Those rows and tarballs are exactly as published on 2026-08-22 and were deliberately left untouched by this snapshot; see MANIFEST_NOTES.md and UNBACKED.tsv.

Reading the results — caveats that change interpretation

  1. method_launch:TimeoutExpired is not a failure. The cell hit the 72h wall-clock backstop; its score is a legitimate anytime result from a truncated search.
  2. budget:max_usd is not a failure either. The cell exhausted its $150 envelope. Two AdaEvolve cells flagged budget-spin (gpqa/strong seed1, seed2) kept iterating after the budget latched — the vendored engine counts a refused evaluation as a completed iteration — so their curve and iteration counts overstate the search that actually happened (~223 of 299 real for seed2; ~190 of 191 for seed1). Their scores are unaffected: nothing after the latch can promote.
  3. rescored rows had their genesis and/or final measured post-hoc. Until 2026-08-29 the driver scanned only the current trace segment for the genesis stamp, so a resumed run took its own resume-time incumbent as genesis and Δ_self collapsed to ~0 by construction. Fixed in mab2.trace.folds.original_genesis; affected cells were re-measured and carry both a rescore block and summary.pre-rescore.json.
  4. Search depth is not uniform. Two OpenAI credit outages (2026-08-27, 2026-08-30) interrupted the AdaEvolve sweep. Interrupted cells were rewound to their last pre-outage checkpoint and resumed; budget now carries across resumes, so caps bind on total spend.
  5. The test split cannot resolve small effects. test_se is 0.035-0.055 (test n=55-66, 3 reps). Detecting a true +0.03 difference needs roughly 4x the test data. Every AdaEvolve Δ_self in this snapshot lies within ~1.6 SE of zero — read that as "not resolvable here", not "measured zero".
  6. invalid non-empty means read the reason first.
Downloads last month
243
Total file size:
9.83 GB