Hugging Face's logo

Datasets:
JG1310
/
repro-flatness-aware-stochastic-bundle

Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
scratch: list<item: struct<optimizer: string, seed: int64, test_acc: double, train_loss: double, s_per_iter:  (... 8 chars omitted)
  child 0, item: struct<optimizer: string, seed: int64, test_acc: double, train_loss: double, s_per_iter: double>
      child 0, optimizer: string
      child 1, seed: int64
      child 2, test_acc: double
      child 3, train_loss: double
      child 4, s_per_iter: double
scratch_summary: struct<SGD: struct<acc_mean: double, acc_std: double, s_per_iter: double>, SAM: struct<acc_mean: dou (... 112 chars omitted)
  child 0, SGD: struct<acc_mean: double, acc_std: double, s_per_iter: double>
      child 0, acc_mean: double
      child 1, acc_std: double
      child 2, s_per_iter: double
  child 1, SAM: struct<acc_mean: double, acc_std: double, s_per_iter: double>
      child 0, acc_mean: double
      child 1, acc_std: double
      child 2, s_per_iter: double
  child 2, fSGLD: struct<acc_mean: double, acc_std: double, s_per_iter: double>
      child 0, acc_mean: double
      child 1, acc_std: double
      child 2, s_per_iter: double
efficiency: struct<sam_over_fsgld_s_per_iter: double, sam_over_sgd: double, fsgld_over_sgd: double>
  child 0, sam_over_fsgld_s_per_iter: double
  child 1, sam_over_sgd: double
  child 2, fsgld_over_sgd: double
ablation: list<item: struct<eta: double, seed: int64, test_acc: double>>
  child 0, item: struct<eta: double, seed: int64, test_acc: double>
      child 0, eta: double
      child 1, seed: int64
      child 2, test
...
beta_grid: list<item: double>
      child 0, item: double
  child 5, beta_residual_D1b: list<item: double>
      child 0, item: double
partB: struct<beta_grid: list<item: double>, KL: list<item: double>, W2: list<item: double>, slope_logKL_lo (... 146 chars omitted)
  child 0, beta_grid: list<item: double>
      child 0, item: double
  child 1, KL: list<item: double>
      child 0, item: double
  child 2, W2: list<item: double>
      child 0, item: double
  child 3, slope_logKL_logbeta: double
  child 4, KL_monotone_decreasing: bool
  child 5, KL_emp_surrogate: double
  child 6, mass_flat_pistar: double
  child 7, mass_flat_expu: double
  child 8, flatness_bias_real: bool
partC: struct<beta: double, lambda_grid: list<item: double>, W1_mean: list<item: double>, W2_mean: list<ite (... 223 chars omitted)
  child 0, beta: double
  child 1, lambda_grid: list<item: double>
      child 0, item: double
  child 2, W1_mean: list<item: double>
      child 0, item: double
  child 3, W2_mean: list<item: double>
      child 0, item: double
  child 4, excess_mean: list<item: double>
      child 0, item: double
  child 5, fit_p_W1: double
  child 6, fit_q_W2: double
  child 7, fit_r_excess: double
  child 8, floor_W1: double
  child 9, floor_W2: double
  child 10, floor_excess: double
  child 11, k_sweep: struct<k: list<item: int64>, W1: list<item: double>>
      child 0, k: list<item: int64>
          child 0, item: int64
      child 1, W1: list<item: double>
          child 0, item: double
to
{'partA': {'slope_D1a': Value('float64'), 'slope_D1b': Value('float64'), 'sigma_grid': List(Value('float64')), 'residual_D1a': List(Value('float64')), 'beta_grid': List(Value('float64')), 'beta_residual_D1b': List(Value('float64'))}, 'partB': {'beta_grid': List(Value('float64')), 'KL': List(Value('float64')), 'W2': List(Value('float64')), 'slope_logKL_logbeta': Value('float64'), 'KL_monotone_decreasing': Value('bool'), 'KL_emp_surrogate': Value('float64'), 'mass_flat_pistar': Value('float64'), 'mass_flat_expu': Value('float64'), 'flatness_bias_real': Value('bool')}, 'partC': {'beta': Value('float64'), 'lambda_grid': List(Value('float64')), 'W1_mean': List(Value('float64')), 'W2_mean': List(Value('float64')), 'excess_mean': List(Value('float64')), 'fit_p_W1': Value('float64'), 'fit_q_W2': Value('float64'), 'fit_r_excess': Value('float64'), 'floor_W1': Value('float64'), 'floor_W2': Value('float64'), 'floor_excess': Value('float64'), 'k_sweep': {'k': List(Value('int64')), 'W1': List(Value('float64'))}}, 'meta': {'eta': Value('float64'), 'n_seeds': Value('int64'), 'lambda_max_est': Value('float64'), 'scale': Value('string')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              scratch: list<item: struct<optimizer: string, seed: int64, test_acc: double, train_loss: double, s_per_iter:  (... 8 chars omitted)
                child 0, item: struct<optimizer: string, seed: int64, test_acc: double, train_loss: double, s_per_iter: double>
                    child 0, optimizer: string
                    child 1, seed: int64
                    child 2, test_acc: double
                    child 3, train_loss: double
                    child 4, s_per_iter: double
              scratch_summary: struct<SGD: struct<acc_mean: double, acc_std: double, s_per_iter: double>, SAM: struct<acc_mean: dou (... 112 chars omitted)
                child 0, SGD: struct<acc_mean: double, acc_std: double, s_per_iter: double>
                    child 0, acc_mean: double
                    child 1, acc_std: double
                    child 2, s_per_iter: double
                child 1, SAM: struct<acc_mean: double, acc_std: double, s_per_iter: double>
                    child 0, acc_mean: double
                    child 1, acc_std: double
                    child 2, s_per_iter: double
                child 2, fSGLD: struct<acc_mean: double, acc_std: double, s_per_iter: double>
                    child 0, acc_mean: double
                    child 1, acc_std: double
                    child 2, s_per_iter: double
              efficiency: struct<sam_over_fsgld_s_per_iter: double, sam_over_sgd: double, fsgld_over_sgd: double>
                child 0, sam_over_fsgld_s_per_iter: double
                child 1, sam_over_sgd: double
                child 2, fsgld_over_sgd: double
              ablation: list<item: struct<eta: double, seed: int64, test_acc: double>>
                child 0, item: struct<eta: double, seed: int64, test_acc: double>
                    child 0, eta: double
                    child 1, seed: int64
                    child 2, test
              ...
              beta_grid: list<item: double>
                    child 0, item: double
                child 5, beta_residual_D1b: list<item: double>
                    child 0, item: double
              partB: struct<beta_grid: list<item: double>, KL: list<item: double>, W2: list<item: double>, slope_logKL_lo (... 146 chars omitted)
                child 0, beta_grid: list<item: double>
                    child 0, item: double
                child 1, KL: list<item: double>
                    child 0, item: double
                child 2, W2: list<item: double>
                    child 0, item: double
                child 3, slope_logKL_logbeta: double
                child 4, KL_monotone_decreasing: bool
                child 5, KL_emp_surrogate: double
                child 6, mass_flat_pistar: double
                child 7, mass_flat_expu: double
                child 8, flatness_bias_real: bool
              partC: struct<beta: double, lambda_grid: list<item: double>, W1_mean: list<item: double>, W2_mean: list<ite (... 223 chars omitted)
                child 0, beta: double
                child 1, lambda_grid: list<item: double>
                    child 0, item: double
                child 2, W1_mean: list<item: double>
                    child 0, item: double
                child 3, W2_mean: list<item: double>
                    child 0, item: double
                child 4, excess_mean: list<item: double>
                    child 0, item: double
                child 5, fit_p_W1: double
                child 6, fit_q_W2: double
                child 7, fit_r_excess: double
                child 8, floor_W1: double
                child 9, floor_W2: double
                child 10, floor_excess: double
                child 11, k_sweep: struct<k: list<item: int64>, W1: list<item: double>>
                    child 0, k: list<item: int64>
                        child 0, item: int64
                    child 1, W1: list<item: double>
                        child 0, item: double
              to
              {'partA': {'slope_D1a': Value('float64'), 'slope_D1b': Value('float64'), 'sigma_grid': List(Value('float64')), 'residual_D1a': List(Value('float64')), 'beta_grid': List(Value('float64')), 'beta_residual_D1b': List(Value('float64'))}, 'partB': {'beta_grid': List(Value('float64')), 'KL': List(Value('float64')), 'W2': List(Value('float64')), 'slope_logKL_logbeta': Value('float64'), 'KL_monotone_decreasing': Value('bool'), 'KL_emp_surrogate': Value('float64'), 'mass_flat_pistar': Value('float64'), 'mass_flat_expu': Value('float64'), 'flatness_bias_real': Value('bool')}, 'partC': {'beta': Value('float64'), 'lambda_grid': List(Value('float64')), 'W1_mean': List(Value('float64')), 'W2_mean': List(Value('float64')), 'excess_mean': List(Value('float64')), 'fit_p_W1': Value('float64'), 'fit_q_W2': Value('float64'), 'fit_r_excess': Value('float64'), 'floor_W1': Value('float64'), 'floor_W2': Value('float64'), 'floor_excess': Value('float64'), 'k_sweep': {'k': List(Value('int64')), 'W1': List(Value('float64'))}}, 'meta': {'eta': Value('float64'), 'n_seeds': Value('int64'), 'lambda_max_est': Value('float64'), 'scale': Value('string')}}
              because column names don't match

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.

Reproduction bundle — Flatness-Aware Stochastic Gradient Langevin Dynamics (fSGLD)

Paper: arXiv 2510.02174 · OpenReview Kw0nhxeTeI · ICML 2026. Logbook: https://huggingface.co/spaces/JG1310/repro-flatness-aware-stochastic-gradient-langevin-dynamics

Contents

  • scripts/exp01_theory.py — theory numerical audit (Claims 1–3), numpy only. Runs in ~6 s.
  • scripts/exp02_neural.py — reduced-scale CIFAR-100N optimizer comparison (Claims 4–6), PyTorch/CPU.
  • specs/ — experiment specifications.
  • results/exp01.json, results/exp02.json — full recorded outputs used for all verdicts.
  • logs/exp01.log, logs/exp02.log — verbatim run logs (Modal 32-core CPU).
  • DERIVATIONS.md — verbatim theorem re-derivations paired with per-step checks.
  • BLOCKERS.md — resource blocker: exp02 is CPU-reduced scale (SmallCNN/0.5M/15k subset/30 ep vs paper's ResNet-34/ViT-B/16, 150/75 ep, A100).
  • poster.html / poster.png / poster_preview.pdf — reproduction poster.

Rerun 

# Theory (Claims 1-3): numpy only, ~6 s, CPU, fully in-scale
python3 scripts/exp01_theory.py            # writes results/exp01.json

# Experiments (Claims 4-6): PyTorch, CPU, reduced scale (~4.5 h on 32 cores)
python3 scripts/exp02_neural.py            # downloads real CIFAR-100 + CIFAR-100N, writes results/exp02.json

Both scripts accept a toy/full toggle inside main(); the recorded results are the FULL (32-core Modal CPU) runs.

Verdicts (see logbook for full evidence)

  • Claim 1 (Prop 3.4): VERIFIED — surrogate residual = 3σ⁴ exact (slope 4.0000), coupling slope −0.1000=−η.
  • Claim 2 (Thm 3.5 / Cor D.16): VERIFIED — W1 p=2.18, W2 q=3.46, exponential-in-k mixing.
  • Claim 3 (Thm 3.8): VERIFIED — excess-risk r=9.97 above finite-β floor.
  • Claim 4 (Table 3): efficiency VERIFIED (SAM 1.92× fSGLD); accuracy NOT reproduced at reduced scale (fSGLD 39.97% < SAM 41.45%). Extraction defect: 62.02% not in paper (true 61.51%).
  • Claim 5 (Table 4): VERIFIED at reduced scale (fSGLD 46.95% > SAM 46.39%, fine-tune regime).
  • Claim 6 (§4.4/§4.5): NOT reproduced at reduced scale — flatness ordering reversed (fSGLD sharper than SGD).
Downloads last month
14
Total file size:
3.29 MB