Hugging Face's logo

Datasets:
jlai300
/
RewardLens-phase2-archive

Tasks:
Visual Question Answering
Text Generation
Tags:
reward-models
multimodal
evidence-dependence
best-of-n
iclr-2027
License:
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
A: double
PFC: double
PSC: double
U2: double
U4: double
U8: double
abstention_rate: double
decoder_family: string
family: string
model_id: string
revision: string
runtime_is_lower_bound: bool
runtime_seconds: double
source_output: string
static_failures: int64
worker: string
bands: struct<attribute: struct<bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: d (... 1119 chars omitted)
  child 0, attribute: struct<bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: in (... 193 chars omitted)
      child 0, bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>, 2.0: str (... 161 chars omitted)
          child 0, 1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
          child 1, 2.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
          child 2, 3.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
      child 1, n_models: int64
  child 1, count: struct<bands: struct<1.0:
...
ars omitted)
      child 0, bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>, 2.0: str (... 161 chars omitted)
          child 0, 1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
          child 1, 2.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
          child 2, 3.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
              child 0, mean_abs_delta_PFC: double
              child 1, mean_abs_delta_PSC: double
              child 2, n_pairs: int64
      child 1, n_models: int64
pairs: list<item: struct<delta_A_pp: double, delta_PFC: double, delta_PSC: double, delta_U8: double, factor (... 65 chars omitted)
  child 0, item: struct<delta_A_pp: double, delta_PFC: double, delta_PSC: double, delta_U8: double, factor: string, m (... 53 chars omitted)
      child 0, delta_A_pp: double
      child 1, delta_PFC: double
      child 2, delta_PSC: double
      child 3, delta_U8: double
      child 4, factor: string
      child 5, model_a: string
      child 6, model_b: string
      child 7, threshold_pp: int64
thresholds_pp: list<item: int64>
  child 0, item: int64
to
{'bands': {'attribute': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'count': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('null'), 'mean_abs_delta_PSC': Value('null'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'presence': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'spatial': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}}, 'pairs': List({'delta_A_pp': Value('float64'), 'delta_PFC': Value('float64'), 'delta_PSC': Value('float64'), 'delta_U8': Value('float64'), 'factor': Value('string'), 'model_a': Value('string'), 'model_b': Value('string'), 'threshold_pp': Value('int64')}), 'thresholds_pp': List(Value('int64'))}
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
              A: double
              PFC: double
              PSC: double
              U2: double
              U4: double
              U8: double
              abstention_rate: double
              decoder_family: string
              family: string
              model_id: string
              revision: string
              runtime_is_lower_bound: bool
              runtime_seconds: double
              source_output: string
              static_failures: int64
              worker: string
              bands: struct<attribute: struct<bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: d (... 1119 chars omitted)
                child 0, attribute: struct<bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: in (... 193 chars omitted)
                    child 0, bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>, 2.0: str (... 161 chars omitted)
                        child 0, 1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                        child 1, 2.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                        child 2, 3.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                    child 1, n_models: int64
                child 1, count: struct<bands: struct<1.0:
              ...
              ars omitted)
                    child 0, bands: struct<1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>, 2.0: str (... 161 chars omitted)
                        child 0, 1.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                        child 1, 2.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                        child 2, 3.0: struct<mean_abs_delta_PFC: double, mean_abs_delta_PSC: double, n_pairs: int64>
                            child 0, mean_abs_delta_PFC: double
                            child 1, mean_abs_delta_PSC: double
                            child 2, n_pairs: int64
                    child 1, n_models: int64
              pairs: list<item: struct<delta_A_pp: double, delta_PFC: double, delta_PSC: double, delta_U8: double, factor (... 65 chars omitted)
                child 0, item: struct<delta_A_pp: double, delta_PFC: double, delta_PSC: double, delta_U8: double, factor: string, m (... 53 chars omitted)
                    child 0, delta_A_pp: double
                    child 1, delta_PFC: double
                    child 2, delta_PSC: double
                    child 3, delta_U8: double
                    child 4, factor: string
                    child 5, model_a: string
                    child 6, model_b: string
                    child 7, threshold_pp: int64
              thresholds_pp: list<item: int64>
                child 0, item: int64
              to
              {'bands': {'attribute': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'count': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('null'), 'mean_abs_delta_PSC': Value('null'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'presence': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}, 'spatial': {'bands': {'1.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '2.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}, '3.0': {'mean_abs_delta_PFC': Value('float64'), 'mean_abs_delta_PSC': Value('float64'), 'n_pairs': Value('int64')}}, 'n_models': Value('int64')}}, 'pairs': List({'delta_A_pp': Value('float64'), 'delta_PFC': Value('float64'), 'delta_PSC': Value('float64'), 'delta_U8': Value('float64'), 'factor': Value('string'), 'model_a': Value('string'), 'model_b': Value('string'), 'threshold_pp': Value('int64')}), 'thresholds_pp': List(Value('int64'))}
              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.

RewardLens Phase II Archive

This is the final clean Hugging Face evidence archive for the completed RewardLens Phase II eight-model experiment.

What this archive contains

  • 8-model experiment evidence
  • static judgments
  • audit judgments
  • Best-of-N pair graphs
  • selections
  • final metrics
  • analysis
  • figures/tables
  • manifests
  • provenance
  • validity metadata and frozen annotation materials where available
  • reproducibility metadata and checksums

Models

  1. Qwen3-VL-4B-Instruct
  2. Gemma-3-4B-it
  3. Molmo-7B-D-0924
  4. Skywork-VL-Reward-7B
  5. Idefics3-8B-Llama3
  6. Phi-3.5-Vision-Instruct
  7. LLaVA-OneVision-Qwen2-7B
  8. InternVL3-8B

Evaluation sizes

Per model:

  • Static: 800
  • Audit: 2400
  • Downstream: 800 pools
  • Pair edges: 22,400

Main scientific result

Within the |Delta A| <= 1pp matched set:

  • qualifying within-factor pairs: 7
  • median |Delta RA|: 15.1pp
  • median |Delta PFC|: 21.0pp

Headline Spatial comparison: Skywork vs Phi-3.5-Vision:

  • Delta A = 0.3pp
  • Delta RA = 38.6pp
  • Delta PFC = 40.5pp

Attribute magnitude-control comparison: Phi-3.5-Vision vs LLaVA-OneVision:

  • Delta A = 0
  • Delta RA = 15.1pp
  • relevant pixel-change = 0.118
  • irrelevant pixel-change = 0.179

Scientific positioning

RewardLens audits selective visual evidence dependence: relevant adaptation plus irrelevant invariance. It is a behavioral diagnostic, not a replacement performance metric.

Downstream result

Downstream Best-of-N utility is retained as a secondary diagnostic. This archive does not claim that PFC or RA predicts downstream utility better than conventional accuracy.

Validity caveats

Spatial relevant edits are larger than irrelevant edits. Attribute is the cleaner magnitude-control case.

The human dual-annotator validity audit sample is frozen, but human labels may still be pending. Orientation-swap robustness may still be pending. This archive does not invent human validity rates or orientation results.

Legacy analysis artifacts may contain historical RQ3 bookkeeping. The final paper drops RQ3 from main claims.

Raw data policy

GQA, COCO, and TallyQA raw payloads are not redistributed. This release provides manifests, source IDs, hashes, and reconstruction/provenance metadata instead.

GitHub

https://github.com/Benjamindaoson/RewardLens

Latest final code line: main includes the final eight-model experiment pipeline and analysis.
Downloads last month
367
Total file size:
149 MB