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.

Snapshot 2026-09-05 · 187 cells.

A cell is one (benchmark x method x seed-agent tier x proposer arm x RNG seed) run. Worker model, budget, evaluator and task splits are held fixed; the method is the variable.

Layout

RESULTS.md      per-cell scores + three comparison metrics
EXCLUDED.md     cells withheld from the primary index, and why
SUPERSEDED.md   cells replaced by a fuller re-run, retained not deleted
MANIFEST.json   sha256 per tarball + table-vs-artifact backing check
MANIFEST_NOTES.md   history of table/artifact drift
cells/live/<arm>_<bench>_<tier>_<method>_seed<N>_run.tar.gz
cells/cro/full_<bench>_<tier>_cro_<arm>_seed<N>_run.tar.gz
cells/live/superseded/<date>/   replaced-but-retained runs
recovery/       codex CLI proposer session logs

Each tarball is a complete run dir:

path	contents
`summary.json`	scores, spend, provenance, validity flags
`summary.pre-rescore.json`	present iff flagged `rescored` — the original, for audit
`trace/*.jsonl`	full event stream: rollouts, evals, method logs, LLM effects
`workspace/`	agent source, candidates, versions/

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
charxiv source images are dropped from tarballs; all archives are credential-scrubbed

Coverage

bench	gepa	mh	cro	adaevolve
gpqa	18	18	16	12
tau2	18	18	18	9
charxiv	18	12	18	12

By proposer arm:

method	sol	opus5
gepa	27	27
mh	27	21
cro	27	25
adaevolve	27	6

Which column to compare on

RESULTS.md carries three score columns because they answer different questions:

column	meaning	caveat
`test`	absolute sealed test score	not comparable across methods that start from different agents
`d_self`	`test - genesis` — each method vs what IT started from	the driver's intended metric; sensitive to WHEN a method first checkpoints
`d_floor`	`test - base` — every method vs one shared no-edit floor	comparable across methods; the fairest single number

d_floor is the recommended headline metric. Every method was handed the same provisioned seed agent, so subtracting the shared ROSTER.json no-edit score compares like with like. d_self is what the driver records, but a method that checkpoints late banks early gains into its genesis and shows a misleadingly small delta.

Caveats that change interpretation

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.
Search depth is not uniform. Infrastructure incidents cost some cells hours, and cells resumed afterwards received a fresh ceiling. Compare within a seed before across seeds.
d_self is - wherever genesis is 0.0000. That is CRO's first-stamp behaviour — the genesis was never scored, so the delta is UNKNOWN, not zero. Use d_floor there.
Cells flagged rescored had their final test evaluation re-run after an outage corrupted the original measurement. The search was untouched; summary.pre-rescore.json preserves the original inside the tarball.
tau2's user simulator is gpt-5.4-mini, not the paper's gpt-4.1. Uniform across all tau2 cells.
25 rows are flagged from-manifest — published on HuggingFace but not reproducible from the local trees that generated this file (the 2026-08-30 AdaEvolve re-run, plus cells retained on HF despite being withheld from the primary index). Their scores come from MANIFEST.json; analysis columns are unknown, not zero.

Integrity

MANIFEST.json records, per row, a sha256 of the cell's tarball and whether the published score is backed by it: the archive's summary.json must agree with the table on test score and termination reason, and any row flagged rescored must additionally carry a rescore block plus summary.pre-rescore.json so the repair is auditable rather than asserted.

Verified 2026-09-05 — 186 rows, 0 unbacked. Every published score is backed by the archive that ships with it.

This check exists because the dump has twice published a table its artifacts did not support: 56 cells on 2026-08-20 and 63 on 2026-08-30, where the tarball carried test_score: 0.0 while the table carried a rescored score. Both times the cause was the same -- the table was regenerated from cluster state while the archives were not. Those 63 rows were re-archived on 2026-09-05 (83 tarballs re-packed and re-uploaded); the verification above confirms the drift is closed.

Two rows are flagged archive-authoritative: the published archive held a fuller re-run than the local tree that generated this table, and the archive won. What a user downloads is the source of truth, not our cluster state.

Regenerate with scripts/archive_manifest.py --dump . --write and gate any publish on --verify. Running it after the fact is how the drift survived twice.

Downloads last month: 373

Total file size:

16 GB