Datasets:

withmartian
/

wayfinder

Error code:   StreamingRowsError
Exception:    TypeError
Message:      Couldn't cast array of type int64 to null
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 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2068, in cast_array_to_feature
                  _c(array.field(name) if name in array_fields else null_array, subfeature)
                  ~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2118, in cast_array_to_feature
                  casted_array_values = _c(array.values, feature.feature)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2152, in cast_array_to_feature
                  return array_cast(
                      array,
                  ...<2 lines>...
                      allow_decimal_to_str=allow_decimal_to_str,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1863, in wrapper
                  return func(array, *args, **kwargs)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2014, in array_cast
                  raise TypeError(f"Couldn't cast array of type {_short_str(array.type)} to {_short_str(pa_type)}")
              TypeError: Couldn't cast array of type int64 to null

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.

Wayfinder

Wayfinder is a research artifact dataset for studying how mechanistic-interpretability-inspired methods affect open language-model behavior across benchmarks.

The dataset stores reproducible experiment outputs from the Wayfinder GitHub repo (https://github.com/withmartian/wayfinder). Each run evaluates a model x benchmark x method(s) x temp configuration and records aggregate scores, benefit metrics, per-sample responses, per-sample scores, provenance, generation settings, and links to the exact prompt set used. Wayfinder treats models, benchmarks, prompt sets, methods, composed methods, temperature, and benefit types as first-class objects so results can be compared across the same prompt populations.

This is not a conventional training dataset. It is a public results and artifact store for analyzing:

accuracy and calibration effects of individual MI-inspired methods
constructive or destructive interference when methods are composed
prompt-set-controlled comparisons across models and benchmarks
model correctness similarity and directional rescue/damage opportunities
method-signal similarity and method diversity
post-run recalibration behavior for composed-method results
model-ensembling opportunities under shared prompt sets

Repository Layout

results/: canonical per-run JSON files. Each file represents one experiment run and includes model, benchmark, method config, generation config, scores, benefit scores, provenance, timing, prompt-set reference, per-sample responses, and per-sample scores.
prompt_sets/: canonical prompt-set artifacts. Result JSONs reference these via prompt_set_ref; prompts are stored here rather than duplicated in every result row.
artifacts/: reusable method assets needed to reproduce or reuse runtime methods (CAA vectors, CLAP probes, MICE calibrators, SEA projections, promoted DLR assets, competence probes, and associated metadata).
composition-recalibration/: sidecar recalibration artifacts for composed-method result JSONs. These do not change the source generations; they provide recalibrated calibration views keyed by source result path and SHA-256.
similarity/: reusable model, benchmark, and method similarity/diversity assets, including model correctness similarity, benchmark/model axis similarity, method-signal similarity, method portfolio scorecards, and dependence/feasibility summaries.
ensemble/: reusable ensemble and cascade-policy assets, including calibrated cascade prototypes and policy/candidate/fold tables.
cell_significance/: per-cell best-method "real-gain" significance verdicts consumed by the results viewer.

Result JSONs

A result JSON is the durable unit for an experiment run. Important fields include:

model_id
benchmark_name
method_configs
composed_methods
gen_config
prompt_set_ref
n_samples
scores
benefit_scores
sample_results
metadata
provenance
timestamp

For modern prompt-set-backed results, per-sample prompt text is not duplicated in sample_results. Consumers should hydrate prompts from the referenced prompt_sets/ artifact when needed.

Prompt Sets

Prompt sets define the exact benchmark slice used by runs. They store shared prompt prefix/suffix text once, plus per-sample prompt bodies, references, sample IDs, and metadata. This keeps result JSONs smaller and makes cross-run comparisons safer because multiple model/method runs can point to the same prompt population.

Intended Uses

This dataset is intended for:

comparing Wayfinder method results across models and benchmarks
analyzing method composition effects
auditing calibration and recalibration behavior
studying model diversity and model-ensembling opportunities
reproducing or extending Wayfinder result-viewer analyses
building secondary analysis assets from shared prompt-set-aligned results

Caveats

Some older n_samples <= 100 results are early experiments and should be treated as hypothesis-generating rather than paper-grade evidence.
Prefer n_samples >= 250 prompt-set-backed results for durable comparisons.
Derived assets such as model similarity, method similarity, and recalibration sidecars are analysis caches over the canonical result JSONs.
Oracle rescue/damage metrics show headroom; they are not deployable routing or ensembling policies by themselves.

Downloads last month: 71

Total file size:

20.8 GB