Dataset Viewer
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
experiment: string
spec_version: string
resolutions: struct<t_range: string, delta: double, baseline_pulls: string>
child 0, t_range: string
child 1, delta: double
child 2, baseline_pulls: string
deviation: string
records: list<item: struct<n: int64, W: int64, memory: int64, mem_frac: double, eps: double, s: int64, algo: (... 374 chars omitted)
child 0, item: struct<n: int64, W: int64, memory: int64, mem_frac: double, eps: double, s: int64, algo: string, rep (... 362 chars omitted)
child 0, n: int64
child 1, W: int64
child 2, memory: int64
child 3, mem_frac: double
child 4, eps: double
child 5, s: int64
child 6, algo: string
child 7, replicates: int64
child 8, err_mean: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 9, err_median: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 10, err_max: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 11, peak_memory: int64
child 12, none_rate: double
child 13, total_pulls: int64
child 14, seeds: list<item: int64>
child 0, item: int64
runtime_seconds: double
job_cores: int64
exp01: string
exp03: string
exp02: string
exp05: string
exp06: string
exp04: string
to
{'exp01': Value('string'), 'exp02': Value('string'), 'exp03': Value('string'), 'exp04': Value('string'), 'exp05': Value('string'), 'exp06': 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
experiment: string
spec_version: string
resolutions: struct<t_range: string, delta: double, baseline_pulls: string>
child 0, t_range: string
child 1, delta: double
child 2, baseline_pulls: string
deviation: string
records: list<item: struct<n: int64, W: int64, memory: int64, mem_frac: double, eps: double, s: int64, algo: (... 374 chars omitted)
child 0, item: struct<n: int64, W: int64, memory: int64, mem_frac: double, eps: double, s: int64, algo: string, rep (... 362 chars omitted)
child 0, n: int64
child 1, W: int64
child 2, memory: int64
child 3, mem_frac: double
child 4, eps: double
child 5, s: int64
child 6, algo: string
child 7, replicates: int64
child 8, err_mean: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 9, err_median: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 10, err_max: struct<mean: double, min: double, max: double, per_run: list<item: double>>
child 0, mean: double
child 1, min: double
child 2, max: double
child 3, per_run: list<item: double>
child 0, item: double
child 11, peak_memory: int64
child 12, none_rate: double
child 13, total_pulls: int64
child 14, seeds: list<item: int64>
child 0, item: int64
runtime_seconds: double
job_cores: int64
exp01: string
exp03: string
exp02: string
exp05: string
exp06: string
exp04: string
to
{'exp01': Value('string'), 'exp02': Value('string'), 'exp03': Value('string'), 'exp04': Value('string'), 'exp05': Value('string'), 'exp06': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Reproduction bundle — Online Learning with Recency: Algorithms for Sliding-window Streaming Multi-armed Bandits
Braverman, Wang, Wang, Zhou. ICML 2026. OpenReview lTqCHLllix, arXiv 2606.08977.
Authors' code: https://github.com/jhwjhw0123/sliding-window-streaming-MABs (independent reimplementation; the authors' code was not used).
Contents
| Path | What it is |
|---|---|
core.py |
Shared model: Instance, Memory, Definition-2 sliding-window semantics, all five instance families, Algorithm 1 (bucket), the top-k baseline, UCB1, the reservoir variants, the exact memory grid |
gates.py |
135 fidelity checks across the six experiments; recomputes the paper's constants independently of the experiment scripts |
scripts/exp01…exp06*.py |
The six experiments (see below) |
specs/ |
Per-experiment specifications written before implementation |
results/*.json |
Full result payloads (2.0 MB) |
results/GATE_REPORT.txt |
Output of python3 gates.py all --full --report |
results/DRIVER_REPORT.json |
Per-experiment pass/fail as recorded by the run driver |
logs/ |
stdout/stderr of every run, including the exp06 crash traceback |
DERIVATIONS.md |
Line-by-line re-derivation of Theorems 1, 2 and 3 with the executable check attached to each step |
BRIEF_WRITER.md |
Claim → evidence map, resolutions, and the pre-registered support/refutation criteria |
BLOCKERS.md |
What failed |
diagnostic_exp06_rootcause.py |
Seconds-scale root-cause diagnostic for the exp06 full-scale crash (output in logs/) |
Experiments
| exp | Purpose | Theorem | Status |
|---|---|---|---|
| exp01 | ε-exploration memory/quality trade-off, synthetic Bern(U(0,1)) | Thm 2 | full scale, gate PASS (15/15) |
| exp02 | ε-exploration on MovieLens 1M | Thm 2 | full scale, gate PASS (30/30) |
| exp03 | Epoch-wise regret vs memory sweep + flat control | Thm 3 / Fig. 10 | full scale, gate PASS (22/22) |
| exp04 | Exact best-arm identification under unlimited pulls | Thm 1 | full scale, gate PASS (21/21) |
| exp05 | Theorem 3 lower bound, upper bound, analytic inequalities | Thm 3 | full scale, gate PASS (19/19) |
| exp06 | PAC guarantee of Algorithm 1 at the paper's exact constants | Thm 2 | full scale BLOCKED; reduced grid, gate 17/18 |
exp06 blocker
python3 scripts/exp06_thm2_pac.py (full scale) dies with
ValueError: p < 0, p > 1 or p is NaN at core.py:69. Root cause, confirmed by
diagnostic_exp06_rootcause.py: the full-scale grid pairs Theorem 1's descending
instance µ_i = 1 − i/(3W) — which the paper states at n = 2W and which is a valid
Bernoulli mean only for i ≤ 3W — with the exp06 stream length n = 5000. At
W = 20 this makes 4940 of 5000 means negative. Only the descending family is
affected; bernoulli_uniform and boundary_adversarial are in range at every
full-scale (n, W). Fix: restrict the descending family to n ≤ 3W, or rescale to
µ_i = 1 − i/(3n). The reduced-grid exp06 results shipped here use
bernoulli_uniform only and are unaffected.
Rerun
Requires Python 3.13, numpy 2.4.6, scipy 1.16.3, joblib 1.5.3. CPU only — no GPU is used
or useful (these are pure numpy bandit simulations).
# MovieLens 1M is downloaded by exp02 into ./data (not shipped here, 30 MB)
JOB_CORES=4 ./run_all.sh # ~2.7 h wall clock on 4 cores
python3 gates.py all --full --report # rewrites results/GATE_REPORT.txt
Every script accepts --toy for a seconds-scale smoke test. All randomness is seeded through
core.seed_for(...), a pure function of the experiment tag and cell parameters, so results are
bit-reproducible.
What is not here
data/ml-1m/ (the MovieLens 1M archive, 30 MB) is not redistributed. exp02 downloads it from
https://files.grouplens.org/datasets/movielens/ml-1m.zip and verifies provenance against the
recorded SHA-256 and against the published row/user/movie counts before running.
- Downloads last month
- 29