Datasets:

sunweiwei
/

ai4sci-eval-results

License:

Dataset Viewer

Split (1)

validation

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
env_version: string
metric_version: string
protocol: string
randomization: string
seeds: list<item: int64>
  child 0, item: int64
scope: string
cases: list<item: struct<case: string, family: string, episodes: list<item: struct<steps: int64, complete:  (... 1372 chars omitted)
  child 0, item: struct<case: string, family: string, episodes: list<item: struct<steps: int64, complete: bool, stead (... 1360 chars omitted)
      child 0, case: string
      child 1, family: string
      child 2, episodes: list<item: struct<steps: int64, complete: bool, steady_start_step: int64, shape_rms_m_full: double,  (... 1232 chars omitted)
          child 0, item: struct<steps: int64, complete: bool, steady_start_step: int64, shape_rms_m_full: double, shape_rms_m (... 1220 chars omitted)
              child 0, steps: int64
              child 1, complete: bool
              child 2, steady_start_step: int64
              child 3, shape_rms_m_full: double
              child 4, shape_rms_m_last_half: double
              child 5, xpoint_error_m_last_half: double
              child 6, ip_error_rel_last_half: double
              child 7, coil_violation_fraction: double
              child 8, max_shape_rms_m: double
              child 9, shape_rms_steady: double
              child 10, xpoint_error_steady: double
              child 11, ip_error_rel_steady: double
              child 12, rin_error_steady: double
              child 13, rout_error_steady: double
              child 14, zx_error_s
...
s: struct<reward: string, success_fraction: string, feasible_fraction: string, tolerance_time_fraction: (... 143 chars omitted)
  child 0, reward: string
  child 1, success_fraction: string
  child 2, feasible_fraction: string
  child 3, tolerance_time_fraction: string
  child 4, shape_rms_cm: string
  child 5, xpoint_error_cm: string
  child 6, rin_error_cm: string
  child 7, rout_error_cm: string
  child 8, zx_error_cm: string
  child 9, ip_error_pct: string
success_tolerances: struct<shape_rms_m: double, gap_m: double, ip_error_rel: double, settle_s: double>
  child 0, shape_rms_m: double
  child 1, gap_m: double
  child 2, ip_error_rel: double
  child 3, settle_s: double
latency: struct<policy_compute_ms: double, per_case_ms: struct<mastu_high_ip_hold: double, mastu_p4_shape_hol (... 105 chars omitted)
  child 0, policy_compute_ms: double
  child 1, per_case_ms: struct<mastu_high_ip_hold: double, mastu_p4_shape_hold: double, mastu_sequence: double, mastu_divert (... 16 chars omitted)
      child 0, mastu_high_ip_hold: double
      child 1, mastu_p4_shape_hold: double
      child 2, mastu_sequence: double
      child 3, mastu_diverted_hold: double
  child 2, steps_per_case: int64
  child 3, limit_ms: double
workers: int64
replay_wall_time_sec: double
wall_time_sec: double
episode_timeout_sec: double
step_silence_sec: double
budget_scales: list<item: double>
  child 0, item: double
max_recovered_steps: int64
n_cases: int64
aggregate_score: double
note: string
n_seeds: int64
to
{'scope': Value('string'), 'randomization': Value('string'), 'n_cases': Value('int64'), 'n_seeds': Value('int64'), 'aggregate_score': Value('float64'), 'cases': {'mastu_high_ip_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_p4_shape_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'coil_limit': Value('int64'), 'kept': Value('int64'), 'plasma_lost': Value('int64'), 'solver_failure': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_sequence': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'solver_failure': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_diverted_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}}, 'note': 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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
              env_version: string
              metric_version: string
              protocol: string
              randomization: string
              seeds: list<item: int64>
                child 0, item: int64
              scope: string
              cases: list<item: struct<case: string, family: string, episodes: list<item: struct<steps: int64, complete:  (... 1372 chars omitted)
                child 0, item: struct<case: string, family: string, episodes: list<item: struct<steps: int64, complete: bool, stead (... 1360 chars omitted)
                    child 0, case: string
                    child 1, family: string
                    child 2, episodes: list<item: struct<steps: int64, complete: bool, steady_start_step: int64, shape_rms_m_full: double,  (... 1232 chars omitted)
                        child 0, item: struct<steps: int64, complete: bool, steady_start_step: int64, shape_rms_m_full: double, shape_rms_m (... 1220 chars omitted)
                            child 0, steps: int64
                            child 1, complete: bool
                            child 2, steady_start_step: int64
                            child 3, shape_rms_m_full: double
                            child 4, shape_rms_m_last_half: double
                            child 5, xpoint_error_m_last_half: double
                            child 6, ip_error_rel_last_half: double
                            child 7, coil_violation_fraction: double
                            child 8, max_shape_rms_m: double
                            child 9, shape_rms_steady: double
                            child 10, xpoint_error_steady: double
                            child 11, ip_error_rel_steady: double
                            child 12, rin_error_steady: double
                            child 13, rout_error_steady: double
                            child 14, zx_error_s
              ...
              s: struct<reward: string, success_fraction: string, feasible_fraction: string, tolerance_time_fraction: (... 143 chars omitted)
                child 0, reward: string
                child 1, success_fraction: string
                child 2, feasible_fraction: string
                child 3, tolerance_time_fraction: string
                child 4, shape_rms_cm: string
                child 5, xpoint_error_cm: string
                child 6, rin_error_cm: string
                child 7, rout_error_cm: string
                child 8, zx_error_cm: string
                child 9, ip_error_pct: string
              success_tolerances: struct<shape_rms_m: double, gap_m: double, ip_error_rel: double, settle_s: double>
                child 0, shape_rms_m: double
                child 1, gap_m: double
                child 2, ip_error_rel: double
                child 3, settle_s: double
              latency: struct<policy_compute_ms: double, per_case_ms: struct<mastu_high_ip_hold: double, mastu_p4_shape_hol (... 105 chars omitted)
                child 0, policy_compute_ms: double
                child 1, per_case_ms: struct<mastu_high_ip_hold: double, mastu_p4_shape_hold: double, mastu_sequence: double, mastu_divert (... 16 chars omitted)
                    child 0, mastu_high_ip_hold: double
                    child 1, mastu_p4_shape_hold: double
                    child 2, mastu_sequence: double
                    child 3, mastu_diverted_hold: double
                child 2, steps_per_case: int64
                child 3, limit_ms: double
              workers: int64
              replay_wall_time_sec: double
              wall_time_sec: double
              episode_timeout_sec: double
              step_silence_sec: double
              budget_scales: list<item: double>
                child 0, item: double
              max_recovered_steps: int64
              n_cases: int64
              aggregate_score: double
              note: string
              n_seeds: int64
              to
              {'scope': Value('string'), 'randomization': Value('string'), 'n_cases': Value('int64'), 'n_seeds': Value('int64'), 'aggregate_score': Value('float64'), 'cases': {'mastu_high_ip_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_p4_shape_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'coil_limit': Value('int64'), 'kept': Value('int64'), 'plasma_lost': Value('int64'), 'solver_failure': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_sequence': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'solver_failure': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}, 'mastu_diverted_hold': {'status': Value('string'), 'n_episodes': Value('int64'), 'n_feasible': Value('int64'), 'n_success': Value('int64'), 'outcomes': {'kept': Value('int64'), 'plasma_lost': Value('int64'), 'success': Value('int64')}, 'fallback_steps': Value('int64'), 'reruns': Value('int64')}}, 'note': 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.

AI4Sci task evaluation results

Final verifier outputs of agent sessions and baselines on the AI4Sci (SSI Index) tasks, October 2026. Each directory is one scored package: reward.json (flat metric breakdown; reward is the task score), and where available grading.json (outcome counts) and results.json (every case or episode). replay-*.json are the same files written by the replay step inside the verifier.

task	run	agent / package	session	verifier	reward
tokamak-magnetic-control	agent-gpt6astra-xhigh-25h-E182_metric-v3	Codex, GPT-6 Astra (xhigh), package E182	25.6 h (stopped at the API quota)	metric v3, 35 sealed episodes	0.667
tokamak-magnetic-control	baseline-shape_vc_sealed_metric-v4_harbor	classical virtual-circuit controller	—	metric v4, 105 sealed episodes, Harbor	0.457
tokamak-magnetic-control	baseline-hold_current_sealed_metric-v4	hold-current anchor	—	metric v4, 105 sealed episodes	0.000
tokamak-magnetic-control	baseline-shape_vc_dev_metric-v4	classical controller	—	metric v4, development suite (21)	0.762
tokamak-magnetic-control	baseline-shape_vc_dev_metric-v3	classical controller	—	metric v3, development suite	0.789
tokamak-magnetic-control	validation-hard-seeds_shape_vc_metric-v4	classical controller	—	metric v4, 28 hardest sealed episodes	0.286
earth-observation-pretraining-transfer	agent-gemini4argon-5h-eo-r2	Antigravity, Gemini 4 Argon	5 h	task verifier	0.406
atomistic-materials	agent-gpt61sol-medium-5h-am-r3	Codex, GPT-6.1 Sol (medium)	5 h	task verifier	0.105
atomistic-materials	reference-mixed-final	reference run	—	task verifier	0.502
protein-sequence-modeling	agent-package-psm-r1	agent package psm-r1	—	task verifier	0.647
fruit-fly-neuron-reconstruction	agent-package-ff-r1_formal-1	agent package (ff-r1)	—	task verifier, formal run	2.2e-5 (nERL)
fluid-flow-control	agent-gpt61sol-medium-5h-fluid-r1c	Codex, GPT-6.1 Sol (medium)	5 h	task verifier	per case (1 failed case)
fluid-flow-control	anchor-random-policy	random policy	—	task verifier	per case (24 failed cases)

Tokamak metric versions: v3 excluded episodes that stalled the solver and measured the latency gate during the replay; v4 counts every episode, gives each solve a deterministic work budget with a bounded fallback ladder, and measures the latency gate after the replay (task evaluation/METRICS.md). The two are not directly comparable.

Downloads last month: 53

Total file size:

705 kB