Datasets:

JG1310
/

repro-nash-linear-bandits-bundle

ArXiv:

Dataset Viewer

Split (1)

train

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
schema_version: int64
skill: string
timestamp: timestamp[s]
poster_html: string
canvas: struct<source: string, width_cm: double, height_cm: double, orientation: string, source_url: null>
  child 0, source: string
  child 1, width_cm: double
  child 2, height_cm: double
  child 3, orientation: string
  child 4, source_url: null
overall: string
hard_failures: int64
warnings: int64
soft_advisories: int64
gates: list<item: struct<name: string, severity: string, status: string, command: list<item: string>, summa (... 229 chars omitted)
  child 0, item: struct<name: string, severity: string, status: string, command: list<item: string>, summary: struct< (... 217 chars omitted)
      child 0, name: string
      child 1, severity: string
      child 2, status: string
      child 3, command: list<item: string>
          child 0, item: string
      child 4, summary: struct<exit_code: int64, tail: string, gate: string, status: string, rules: list<item: struct<id: in (... 73 chars omitted)
          child 0, exit_code: int64
          child 1, tail: string
          child 2, gate: string
          child 3, status: string
          child 4, rules: list<item: struct<id: int64, severity: string, status: string, detail: string>>
              child 0, item: struct<id: int64, severity: string, status: string, detail: string>
                  child 0, id: int64
                  child 1, severity: string
                  child 2, status: string
                  child 3, detail: string
      
...
d: bool
      child 6, frac_of_T_spent_in_phase1: double
envelope_summary: struct<n_cells: int64, violations_with_flat_C16: int64, violations_with_per_algo_C: int64, residual: (... 114 chars omitted)
  child 0, n_cells: int64
  child 1, violations_with_flat_C16: int64
  child 2, violations_with_per_algo_C: int64
  child 3, residual: list<item: struct<algo: string, d: int64, T: int64, nr: double, bound_correct_C: double, ratio_corre (... 12 chars omitted)
      child 0, item: struct<algo: string, d: int64, T: int64, nr: double, bound_correct_C: double, ratio_correct: double>
          child 0, algo: string
          child 1, d: int64
          child 2, T: int64
          child 3, nr: double
          child 4, bound_correct_C: double
          child 5, ratio_correct: double
head_to_head_summary: struct<comparisons: int64, fairlin_wins: int64, losses: int64, max_ratio: double, min_ratio: double>
  child 0, comparisons: int64
  child 1, fairlin_wins: int64
  child 2, losses: int64
  child 3, max_ratio: double
  child 4, min_ratio: double
T_exponent_by_regime: list<item: struct<d: int64, algo: string, slope_all_T: double, n_phase2_cells: int64, slope_phase2_o (... 13 chars omitted)
  child 0, item: struct<d: int64, algo: string, slope_all_T: double, n_phase2_cells: int64, slope_phase2_only: double (... 1 chars omitted)
      child 0, d: int64
      child 1, algo: string
      child 2, slope_all_T: double
      child 3, n_phase2_cells: int64
      child 4, slope_phase2_only: double
to
{'phase1_regime': List({'d': Value('int64'), 'T': Value('int64'), 'mu_star': Value('float64'), 'tau_lo': Value('float64'), 'tau_hi': Value('float64'), 'phase2_reached': Value('bool'), 'frac_of_T_spent_in_phase1': Value('float64')}), 'envelope': List({'algo': Value('string'), 'd': Value('int64'), 'T': Value('int64'), 'nr': Value('float64'), 'bound_C16': Value('float64'), 'holds_C16': Value('bool'), 'bound_correct_C': Value('float64'), 'correct_C': Value('float64'), 'holds_correct_C': Value('bool'), 'ratio_correct': Value('float64')}), 'envelope_summary': {'n_cells': Value('int64'), 'violations_with_flat_C16': Value('int64'), 'violations_with_per_algo_C': Value('int64'), 'residual': List({'algo': Value('string'), 'd': Value('int64'), 'T': Value('int64'), 'nr': Value('float64'), 'bound_correct_C': Value('float64'), 'ratio_correct': Value('float64')})}, 'head_to_head': List({'d': Value('int64'), 'T': Value('int64'), 'lin_nash': Value('float64'), 'fair_lin_pe': Value('float64'), 'fair_lin_pe_wins': Value('bool'), 'fair_lin_pe_ratio': Value('float64'), 'fair_lin_ucb': Value('float64'), 'fair_lin_ucb_wins': Value('bool'), 'fair_lin_ucb_ratio': Value('float64')}), 'head_to_head_summary': {'comparisons': Value('int64'), 'fairlin_wins': Value('int64'), 'losses': Value('int64'), 'max_ratio': Value('float64'), 'min_ratio': Value('float64')}, 'T_exponent_by_regime': List({'d': Value('int64'), 'algo': Value('string'), 'slope_all_T': Value('float64'), 'n_phase2_cells': Value('int64'), 'slope_phase2_only': Value('float64')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 478, 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
              schema_version: int64
              skill: string
              timestamp: timestamp[s]
              poster_html: string
              canvas: struct<source: string, width_cm: double, height_cm: double, orientation: string, source_url: null>
                child 0, source: string
                child 1, width_cm: double
                child 2, height_cm: double
                child 3, orientation: string
                child 4, source_url: null
              overall: string
              hard_failures: int64
              warnings: int64
              soft_advisories: int64
              gates: list<item: struct<name: string, severity: string, status: string, command: list<item: string>, summa (... 229 chars omitted)
                child 0, item: struct<name: string, severity: string, status: string, command: list<item: string>, summary: struct< (... 217 chars omitted)
                    child 0, name: string
                    child 1, severity: string
                    child 2, status: string
                    child 3, command: list<item: string>
                        child 0, item: string
                    child 4, summary: struct<exit_code: int64, tail: string, gate: string, status: string, rules: list<item: struct<id: in (... 73 chars omitted)
                        child 0, exit_code: int64
                        child 1, tail: string
                        child 2, gate: string
                        child 3, status: string
                        child 4, rules: list<item: struct<id: int64, severity: string, status: string, detail: string>>
                            child 0, item: struct<id: int64, severity: string, status: string, detail: string>
                                child 0, id: int64
                                child 1, severity: string
                                child 2, status: string
                                child 3, detail: string
                    
              ...
              d: bool
                    child 6, frac_of_T_spent_in_phase1: double
              envelope_summary: struct<n_cells: int64, violations_with_flat_C16: int64, violations_with_per_algo_C: int64, residual: (... 114 chars omitted)
                child 0, n_cells: int64
                child 1, violations_with_flat_C16: int64
                child 2, violations_with_per_algo_C: int64
                child 3, residual: list<item: struct<algo: string, d: int64, T: int64, nr: double, bound_correct_C: double, ratio_corre (... 12 chars omitted)
                    child 0, item: struct<algo: string, d: int64, T: int64, nr: double, bound_correct_C: double, ratio_correct: double>
                        child 0, algo: string
                        child 1, d: int64
                        child 2, T: int64
                        child 3, nr: double
                        child 4, bound_correct_C: double
                        child 5, ratio_correct: double
              head_to_head_summary: struct<comparisons: int64, fairlin_wins: int64, losses: int64, max_ratio: double, min_ratio: double>
                child 0, comparisons: int64
                child 1, fairlin_wins: int64
                child 2, losses: int64
                child 3, max_ratio: double
                child 4, min_ratio: double
              T_exponent_by_regime: list<item: struct<d: int64, algo: string, slope_all_T: double, n_phase2_cells: int64, slope_phase2_o (... 13 chars omitted)
                child 0, item: struct<d: int64, algo: string, slope_all_T: double, n_phase2_cells: int64, slope_phase2_only: double (... 1 chars omitted)
                    child 0, d: int64
                    child 1, algo: string
                    child 2, slope_all_T: double
                    child 3, n_phase2_cells: int64
                    child 4, slope_phase2_only: double
              to
              {'phase1_regime': List({'d': Value('int64'), 'T': Value('int64'), 'mu_star': Value('float64'), 'tau_lo': Value('float64'), 'tau_hi': Value('float64'), 'phase2_reached': Value('bool'), 'frac_of_T_spent_in_phase1': Value('float64')}), 'envelope': List({'algo': Value('string'), 'd': Value('int64'), 'T': Value('int64'), 'nr': Value('float64'), 'bound_C16': Value('float64'), 'holds_C16': Value('bool'), 'bound_correct_C': Value('float64'), 'correct_C': Value('float64'), 'holds_correct_C': Value('bool'), 'ratio_correct': Value('float64')}), 'envelope_summary': {'n_cells': Value('int64'), 'violations_with_flat_C16': Value('int64'), 'violations_with_per_algo_C': Value('int64'), 'residual': List({'algo': Value('string'), 'd': Value('int64'), 'T': Value('int64'), 'nr': Value('float64'), 'bound_correct_C': Value('float64'), 'ratio_correct': Value('float64')})}, 'head_to_head': List({'d': Value('int64'), 'T': Value('int64'), 'lin_nash': Value('float64'), 'fair_lin_pe': Value('float64'), 'fair_lin_pe_wins': Value('bool'), 'fair_lin_pe_ratio': Value('float64'), 'fair_lin_ucb': Value('float64'), 'fair_lin_ucb_wins': Value('bool'), 'fair_lin_ucb_ratio': Value('float64')}), 'head_to_head_summary': {'comparisons': Value('int64'), 'fairlin_wins': Value('int64'), 'losses': Value('int64'), 'max_ratio': Value('float64'), 'min_ratio': Value('float64')}, 'T_exponent_by_regime': List({'d': Value('int64'), 'algo': Value('string'), 'slope_all_T': Value('float64'), 'n_phase2_cells': Value('int64'), 'slope_phase2_only': Value('float64')})}
              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 — Improved Algorithms for Nash Welfare in Linear Bandits

Sarkar, Pandey & Ray Chowdhury. arXiv 2601.22969 · OpenReview 5WVIbxWqwA · ICML 2026. Authors' code: https://github.com/NP-Hardest/FairLinBandit

Logbook: https://huggingface.co/spaces/JG1310/repro-improved-algorithms-for-nash-welfare-in-linear-bandits

Verdicts

Claim	Paper result	Verdict
1 — order-optimal Nash regret, resolving the `d` suboptimality	Theorem 4.1 (not "Theorem 66")	VERIFIED — derivation reproduces; envelope holds 39/40 cells; predicted `T^{-1/2}` decay confirmed in-regime. `d^1` vs `d^{1.25}` not measurable on LTR instances.
2 — sublinear p-means regret for the entire range of `p`	Theorem 4.2 (not "Theorem 70")	VERIFIED as a theorem — `p ≥ 0` branch follows from verified components; monotonicity in `p` confirmed. Sublinearity sweep inconclusive (out of regime).
3 — consistently outperforms SOTA on real-world instances	§6, Figs 1–3, Table 1	VERIFIED on MSLR-WEB10K — 19/20 full-scale cells, one 0.5% loss. Yahoo! LTRC blocked. Table 1's runtime ordering does not reproduce.

The headline finding

The paper's own d = 10 MSLR instance has μ* = 0.0667. Theorem 4.1's Phase-I length is then τ ∈ [6.7e7, 2.4e8], and its explicit envelope stays above μ* — vacuous — until T* = 2.7e7. Below T* the algorithm is in pure exploration and regret is flat; above it, the predicted T^{-1/2} decay appears. Computing T*(d) from the appendix constants and μ* alone separates the decaying cells from the flat ones exactly (scripts/diag_tslope.py). The paper's headline horizon T = 10⁸ is the first decade at which its own theorem is non-vacuous on its own instance. The appendix's stated condition T ≥ Ω̃(σ²d²) understates this by ~5 orders of magnitude, because the binding quantity is μ*, not σ and d.

Layout

scripts/    exp00–exp04 (built from specs/) + diag_*.py post-hoc analyses + figure builders
specs/      per-experiment specifications written before implementation
core/       shared modules: MSLR loader, instance builder, environment, geometry, 3 algorithms
results/    exp00–exp04 JSON + diag_*.json + GATE_REPORT.txt + DRIVER_REPORT.json
logs/       stdout of every experiment run
figures/    plotly HTML + raw CSV for each logbook figure
poster/     the reproduction poster (posterly): poster.html, poster_embed.html (the pinned
            logbook figure cell), poster_preview.pdf/.png, images/, GATE_REPORT.json
DERIVATIONS.md   13-step re-derivation of Thm 4.1 (D1–D13) and Thm 4.2 (E1–E4), each paired
                 with the numerical check that tests it
BRIEF_WRITER.md  claim -> evidence map with pre-registered "settles it either way" criteria
BLOCKERS.md      Yahoo! LTRC (unobtainable) and the exp02/exp03 run failures
STATE.md         planning ledger, deviations, theorem-numbering note
gates.py         fidelity gates (procedure only — no gate encodes an expected outcome)
pipeline.log     full driver log incl. the OOM kills and repair attempts
paper.pdf        the paper

Reproduce

pip install numpy scipy scikit-learn plotly
# MSLR-WEB10K (Fold1) must be present; see core/ for the expected layout.
./run_all.sh                       # exp00 -> exp04, ~4.5 h at 12 cores, ~8 GB peak RSS
python3 gates.py                   # regenerates results/GATE_REPORT.txt

Post-hoc analyses — seconds, need only results/*.json (and work/*.npy for the salvage):

python3 scripts/diag_regime.py          # T*(d): where Thm 4.1 stops being vacuous
python3 scripts/diag_tslope.py          # does T* predict which cells decay?  -> yes
python3 scripts/diag_headtohead.py      # Claim 3: 19/20 cells, + LinNash instability
python3 scripts/diag_pmeans_salvage.py  # Claim 2: p-grid recovered from exp03's crash
python3 scripts/make_logbook_figures.py # figures/*.html + *.csv

Known limitations

Yahoo! LTRC blocked — Webscope account + signed data-use agreement required; webscope.sandbox.yahoo.com unreachable, no mirror. A synthetic surrogate was forbidden by spec and not substituted. The paper's d-ablation ran on Yahoo; ours runs on MSLR.
exp02 and exp03 failed at full scale — OOM-killed (8+ GB vs a declared 5 GB budget) across four attempts each; the repair agent then hit a session limit. Their JSON retains 2-replicate T = 10⁴ smoke configs. The sublinearity entries in exp03.json (slope: 0.0, r2: 1.0) are degenerate single-point fits and are not evidence.
T = 10⁸ never reached. The paper's headline horizon — and, per the finding above, the only horizon at which d = 10 is in regime — is outside our compute budget.
work/*.f32.npy (1.4 GB) is not included in this bundle. These are the 48 surviving T = 10⁷ per-round expected-reward traces that diag_pmeans_salvage.py consumes; results/exp03_salvage.json holds its full output.
Gates verify procedure fidelity only — schema, shapes, ranges, dataset provenance. A green gate means "the experiment ran as specified", never "the paper was confirmed". gates.py also mis-scores the Thm 4.1 envelope: it applies C = 16 to LinPE (whose constant is 36, Lemma B.20) and to LinNash (which the theorem does not cover).

Downloads last month: 45

Total file size:

9.04 MB

Paper for JG1310/repro-nash-linear-bandits-bundle

Improved Algorithms for Nash Welfare in Linear Bandits

Paper • 2601.22969 • Published Jun 21