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/folder_based_builder/folder_based_builder.py", line 246, in _split_generators
                  raise ValueError(
                      "`file_name`, `*_file_name`, `file_names` or `*_file_names` must be present as dictionary key in metadata files"
                  )
              ValueError: `file_name`, `*_file_name`, `file_names` or `*_file_names` must be present as dictionary key in metadata files
              
              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.

Hearing a metric get it wrong

Audio examples from the ICASSP 2027 submission "EITL: Ear-in-the-Loop Music Mixing Measures and Mitigates Reward Hacking of Learned Quality Metrics." Anonymised for double-anonymous review. Published for research purposes: these clips are the evidence behind a claim in the paper, and the claim is hard to believe without hearing them.

A learned audio-quality metric is good enough to rank mixes. Point a search at it, accepting any edit that raises the score, and the score keeps climbing past the professional mix while the audio gets worse. These clips let you hear that happen.

The budget ladder

Files named SDRNR_ladder_* follow one song as the search budget k grows. Every clip is the same 12-second passage at matched loudness, so the only difference is how long the search ran. The reward is the mean of two learned metrics.

file metric says (PQ, 1-10) listeners said (0-100)
SDRNR_ladder_k00000_corrupted-start 6.82 27.1
SDRNR_ladder_k00028_search 7.68 54.3
SDRNR_ladder_k00100_search 8.05 40.2
SDRNR_ladder_k01000_search 8.24 26.9
SDRNR_ladder_k04000_search 8.31 23.8
SDRNR_ladder_professional-mix 7.19 60.3
SDRNR_ladder_anchor-lowpass3.5kHz 5.34 10.5

The metric rises the whole way and rates every render from k=100 onward above the professional mix. Listeners peak early and then fall off a cliff: by k=4000 they rate the render below the corrupted session the search started from. Play k04000 against professional-mix and the disagreement is not subtle.

Listener figures are means over the 12 people who rated this song in the crowdsourced study (0-100 sliders, headphone-screened, loudness-matched excerpts). They are from an interim sample and will shift slightly in the final paper; the ordering is what matters here.

Reward forms at the deployment budget

Files named *_k00028_* hold two songs under all eight conditions compared in the paper, at the budget we actually deploy: the professional mix, the corrupted starting session, two external systems (baseline-FxNorm, baseline-MEGAMI), and four reward forms - the minimum of the two calibrated metrics (EITL-min-ours), their weighted mean (EITL-mean), and each metric alone (reward-PQ-only, reward-SB-only).

reward-SB-only is the one to listen for. Optimising the mixing metric alone lets the unmonitored production-quality axis fall furthest, and it was rated lowest of the four reward forms. The min gate is what prevents that trade.

Files

Every clip is provided twice: wav/ (24-bit PCM) and opus/ (48 kHz, 193 kbps). The Opus files are what the listening study actually served, so they are the stimuli of record that the ratings correspond to.

metadata.csv lists each clip with its song, set, condition, search budget, and an audio_source column:

  • source-wav (17 clips): rendered from the uncompressed master with the same excerpt window, gain, and any transform the study applied, and verified against the served Opus by waveform correlation (r > 0.95). These have never been through a lossy codec.
  • decoded-from-served-opus (6 clips): the served Opus decoded to WAV. For these conditions the uncompressed master is no longer available, so rather than substitute a similar-sounding render we ship exactly what the listeners heard.

Source material and licence

The two songs are The Easton Ellises - "Falcon 69" and The Easton Ellises (Baumi) - "SDRNR", from MUSDB18-HQ. We use only these two because they carry an explicit Creative Commons licence in that corpus; MUSDB18's other tracks have mixed provenance and are distributed for educational use under terms that do not clearly permit redistributing derivatives.

These clips are derivative works: the stems were gain-corrupted and then re-processed by the search. They are released for research purposes as supporting material for a scientific publication, under CC BY-NC-SA 4.0, matching the source terms. Attribute The Easton Ellises, non-commercial use only, and share any adaptation alike. If you hold rights in this material and would like it removed, please open a discussion on this dataset.

Downloads last month
49