Datasets:

khazic
/

randkv-kv-scaleup

License:

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
args: struct<bank_size: int64, candidate_parallel: int64, d1_per_task: int64, enforce_eager: bool, gpu_mem (... 403 chars omitted)
  child 0, bank_size: int64
  child 1, candidate_parallel: int64
  child 2, d1_per_task: int64
  child 3, enforce_eager: bool
  child 4, gpu_memory_utilization: double
  child 5, manifest_root: string
  child 6, max_model_len: int64
  child 7, max_new_tokens_cap: null
  child 8, max_num_batched_tokens: int64
  child 9, max_num_seqs: int64
  child 10, model: string
  child 11, output: string
  child 12, prefix_length: int64
  child 13, randopt_root: string
  child 14, repeat_seed: int64
  child 15, samplers: list<item: string>
      child 0, item: string
  child 16, screen_keep: int64
  child 17, screen_per_task: int64
  child 18, smoke: bool
  child 19, smoke_max_new_tokens: int64
  child 20, tasks: list<item: string>
      child 0, item: string
  child 21, wikitext: string
candidate_count: int64
corpus_sha256: string
d1_row_ids: struct<countdown: list<item: int64>, gsm8k: list<item: int64>, math500: list<item: int64>, mbpp: lis (... 109 chars omitted)
  child 0, countdown: list<item: int64>
      child 0, item: int64
  child 1, gsm8k: list<item: int64>
      child 0, item: int64
  child 2, math500: list<item: int64>
      child 0, item: int64
  child 3, mbpp: list<item: int64>
      child 0, item: int64
  child 4, olympiadbench: list<item: int64>
      child 0, item: int64
  child 5, rocstories: list<item: int64>
      child 0, item: int64
  child 6, uspto50k: list<item: int64>
      child 0, item: int64
git_commit: string
gpu_uuid: string
split_manifest_sha256: struct<countdown: string, gsm8k: string, math500: string, mbpp: string, olympiadbench: string, rocst (... 32 chars omitted)
  child 0, countdown: string
  child 1, gsm8k: string
  child 2, math500: string
  child 3, mbpp: string
  child 4, olympiadbench: string
  child 5, rocstories: string
  child 6, uspto50k: string
started_unix: double
prefix_text: string
arm: string
tokenizer: string
repeat_seed: int64
length: int64
prefix_ids: list<item: int64>
  child 0, item: int64
to
{'arm': Value('string'), 'corpus_sha256': Value('string'), 'length': Value('int64'), 'prefix_ids': List(Value('int64')), 'prefix_text': Value('string'), 'repeat_seed': Value('int64'), 'tokenizer': Value('string')}
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 483, 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              args: struct<bank_size: int64, candidate_parallel: int64, d1_per_task: int64, enforce_eager: bool, gpu_mem (... 403 chars omitted)
                child 0, bank_size: int64
                child 1, candidate_parallel: int64
                child 2, d1_per_task: int64
                child 3, enforce_eager: bool
                child 4, gpu_memory_utilization: double
                child 5, manifest_root: string
                child 6, max_model_len: int64
                child 7, max_new_tokens_cap: null
                child 8, max_num_batched_tokens: int64
                child 9, max_num_seqs: int64
                child 10, model: string
                child 11, output: string
                child 12, prefix_length: int64
                child 13, randopt_root: string
                child 14, repeat_seed: int64
                child 15, samplers: list<item: string>
                    child 0, item: string
                child 16, screen_keep: int64
                child 17, screen_per_task: int64
                child 18, smoke: bool
                child 19, smoke_max_new_tokens: int64
                child 20, tasks: list<item: string>
                    child 0, item: string
                child 21, wikitext: string
              candidate_count: int64
              corpus_sha256: string
              d1_row_ids: struct<countdown: list<item: int64>, gsm8k: list<item: int64>, math500: list<item: int64>, mbpp: lis (... 109 chars omitted)
                child 0, countdown: list<item: int64>
                    child 0, item: int64
                child 1, gsm8k: list<item: int64>
                    child 0, item: int64
                child 2, math500: list<item: int64>
                    child 0, item: int64
                child 3, mbpp: list<item: int64>
                    child 0, item: int64
                child 4, olympiadbench: list<item: int64>
                    child 0, item: int64
                child 5, rocstories: list<item: int64>
                    child 0, item: int64
                child 6, uspto50k: list<item: int64>
                    child 0, item: int64
              git_commit: string
              gpu_uuid: string
              split_manifest_sha256: struct<countdown: string, gsm8k: string, math500: string, mbpp: string, olympiadbench: string, rocst (... 32 chars omitted)
                child 0, countdown: string
                child 1, gsm8k: string
                child 2, math500: string
                child 3, mbpp: string
                child 4, olympiadbench: string
                child 5, rocstories: string
                child 6, uspto50k: string
              started_unix: double
              prefix_text: string
              arm: string
              tokenizer: string
              repeat_seed: int64
              length: int64
              prefix_ids: list<item: int64>
                child 0, item: int64
              to
              {'arm': Value('string'), 'corpus_sha256': Value('string'), 'length': Value('int64'), 'prefix_ids': List(Value('int64')), 'prefix_text': Value('string'), 'repeat_seed': Value('int64'), 'tokenizer': 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.

RandKV kv-scaleup campaign results

Raw outputs of the RandKV prefix-local KV-noise scale-up campaign (https://github.com/xxraincandyxx/RandKV, scorers/handlers from https://github.com/sunrainyg/RandOpt).

Two models (Qwen2.5-3B-Instruct, Qwen3-8B) x seven tasks (countdown, gsm8k, math500, mbpp, olympiadbench, rocstories, uspto50k) x two prefix arms (selected = frozen Stage 1 prefix, random = registered random span) x three repeat seeds (42, 101, 202) = 84 Stage 2 cells, plus 6 Stage 1 prefix selections. Every cell: 1024 candidates, top-k 64, vLLM backend.

Layout

summary.csv / summary.json: one row per Stage 2 cell, outcome.json flattened (raw / p0 / rank1 / ensemble accuracy, deltas with bootstrap CIs, wall time, ...).
stage1/<model>/seed-<s>/: the selected prefix (prefix.json) with its manifest and selection scores; stage1/<model>/prefixes/: random-arm span registries.
raw/<model>/<task>.tar.gz: full per-cell directories (<model>/<task>/kv/<arm>/seed-<s>/ with d2/, test/, ensemble.json, selection.json, manifest.json, rms.json, outcome.json). raw/<model>/stage1.tar.gz: Stage 1 including all candidate generations. Failed-attempt leftovers were excluded.

Mean accuracy over seeds

model	task	arm	n	raw	p0	rank1	ensemble
Qwen2.5-3B-Instruct	countdown	random	3	0.101	0.088	0.143	0.506
Qwen2.5-3B-Instruct	countdown	selected	3	0.101	0.098	0.128	0.523
Qwen2.5-3B-Instruct	gsm8k	random	3	0.818	0.753	0.792	0.819
Qwen2.5-3B-Instruct	gsm8k	selected	3	0.818	0.764	0.812	0.827
Qwen2.5-3B-Instruct	math500	random	3	0.602	0.571	0.583	0.627
Qwen2.5-3B-Instruct	math500	selected	3	0.602	0.574	0.574	0.630
Qwen2.5-3B-Instruct	mbpp	random	3	0.591	0.577	0.580	0.608
Qwen2.5-3B-Instruct	mbpp	selected	3	0.591	0.578	0.579	0.580
Qwen2.5-3B-Instruct	olympiadbench	random	3	0.283	0.318	0.314	0.396
Qwen2.5-3B-Instruct	olympiadbench	selected	3	0.283	0.274	0.305	0.343
Qwen2.5-3B-Instruct	rocstories	random	3	0.545	0.541	0.545	0.547
Qwen2.5-3B-Instruct	rocstories	selected	3	0.545	0.540	0.542	0.547
Qwen2.5-3B-Instruct	uspto50k	random	3	0.060	0.077	0.117	0.102
Qwen2.5-3B-Instruct	uspto50k	selected	3	0.060	0.116	0.155	0.135
Qwen3-8B	countdown	random	3	0.557	0.522	0.520	0.738
Qwen3-8B	countdown	selected	3	0.557	0.511	0.511	0.732
Qwen3-8B	gsm8k	random	3	0.864	0.819	0.842	0.857
Qwen3-8B	gsm8k	selected	3	0.864	0.822	0.826	0.853
Qwen3-8B	math500	random	3	0.759	0.750	0.744	0.802
Qwen3-8B	math500	selected	3	0.759	0.735	0.741	0.809
Qwen3-8B	mbpp	random	3	0.742	0.726	0.718	0.735
Qwen3-8B	mbpp	selected	3	0.742	0.705	0.711	0.714
Qwen3-8B	olympiadbench	random	3	0.538	0.500	0.531	0.582
Qwen3-8B	olympiadbench	selected	3	0.538	0.503	0.525	0.563
Qwen3-8B	rocstories	random	3	0.719	0.703	0.710	0.710
Qwen3-8B	rocstories	selected	3	0.719	0.697	0.705	0.704
Qwen3-8B	uspto50k	random	3	0.116	0.083	0.109	0.185
Qwen3-8B	uspto50k	selected	3	0.116	0.111	0.120	0.201

Provenance

RandKV commit 6ed7a5ab530227c36001a00ace0c0d468b43e10a, RandOpt commit 536df0a308f3990b6270c991fbb96bd0b779a58e
vllm / transformers / torch: 0.10.2 4.53.3 2.8.0+cu128
packed 2026-09-04 16:59:35

Downloads last month: 22

Total file size:

737 MB