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
task_count: int64
tasks_4h: int64
tasks_8h: int64
model_calls: int64
codex_sessions: int64
model: string
all_tasks_exit_zero: bool
score_entries_per_task: struct<arc_compiler_runtime: int64, exchange_core_throughput: int64, integer_compression_codec: int6 (... 1521 chars omitted)
  child 0, arc_compiler_runtime: int64
  child 1, exchange_core_throughput: int64
  child 2, integer_compression_codec: int64
  child 3, vliw_kernel_optimization: int64
  child 4, apple_incremental_game: int64
  child 5, grid_turing_robot: int64
  child 6, molecular_self_assembly: int64
  child 7, order_addition_permutation_optimization: int64
  child 8, treant_forest: int64
  child 9, bipedalwalker_locomotion_rl: int64
  child 10, dabic_gravity_inversion: int64
  child 11, ffmpeg_swscale_reimplementation: int64
  child 12, schemathesis_config_modernization: int64
  child 13, schemathesis_reporting_observability: int64
  child 14, borden_source_inversion: int64
  child 15, juliet_vulnerability_analyzer: int64
  child 16, schemathesis_datagen_pipeline: int64
  child 17, ad_placement_optimization: int64
  child 18, jagua_nesting_optimization: int64
  child 19, git_rewrite_in_zig: int64
  child 20, rust_multicrate_reconstruction: int64
  child 21, equivalence_class_divide_and_conquer: int64
  child 22, graph_node_classification: int64
  child 23, ann_vector_search_qps: int64
  child 24, tree_block_partitioning: int64
  child 25, triangulation_coloring_optimization: int64
  child 26, vehicle_routing_time_windows: int64
  child 27, warehouse_forklift_routing: int64
  child 28, wireless_electricity_layout: int64
  child 29, ordinal_notation_well_foundedness: int64
  child 30, anchorhead_text_adventure: int64
  child 31, dcss_dungeon_ai: int64
  child 32, openttd_transport_ai: int64
  child 33, tryst_text_adventure: int64
  child 34, wesnoth_tactical_ai: int64
  child 35, vibrating_path_graph_coloring: int64
  child 36, cta_risk_budget_optimization: int64
  child 37, k12_math_recommendation: int64
  child 38, portfolio_risk_calibration: int64
  child 39, combinatorial_games_formalization: int64
  child 40, new_foundations_consistency: int64
  child 41, pfr_formalization: int64
  child 42, sphere_eversion_formalization: int64
  child 43, trinity_text_adventure: int64
  child 44, nethack_dungeon_agent: int64
  child 45, flt_regular_formalization: int64
  child 46, openrct2_theme_park_ai: int64
zip_bytes: int64
shortened_text_fields: int64
zip_count: int64
tasks: int64
score_entries: int64
trajectories_unchanged: bool
max_allowed_zip_count: int64
to
{'tasks': Value('int64'), 'model_calls': Value('int64'), 'tasks_4h': Value('int64'), 'tasks_8h': Value('int64'), 'score_entries': Value('int64'), 'score_entries_per_task': {'arc_compiler_runtime': Value('int64'), 'exchange_core_throughput': Value('int64'), 'integer_compression_codec': Value('int64'), 'vliw_kernel_optimization': Value('int64'), 'apple_incremental_game': Value('int64'), 'grid_turing_robot': Value('int64'), 'molecular_self_assembly': Value('int64'), 'order_addition_permutation_optimization': Value('int64'), 'treant_forest': Value('int64'), 'bipedalwalker_locomotion_rl': Value('int64'), 'dabic_gravity_inversion': Value('int64'), 'ffmpeg_swscale_reimplementation': Value('int64'), 'schemathesis_config_modernization': Value('int64'), 'schemathesis_reporting_observability': Value('int64'), 'borden_source_inversion': Value('int64'), 'juliet_vulnerability_analyzer': Value('int64'), 'schemathesis_datagen_pipeline': Value('int64'), 'ad_placement_optimization': Value('int64'), 'jagua_nesting_optimization': Value('int64'), 'git_rewrite_in_zig': Value('int64'), 'rust_multicrate_reconstruction': Value('int64'), 'equivalence_class_divide_and_conquer': Value('int64'), 'graph_node_classification': Value('int64'), 'ann_vector_search_qps': Value('int64'), 'tree_block_partitioning': Value('int64'), 'triangulation_coloring_optimization': Value('int64'), 'vehicle_routing_time_windows': Value('int64'), 'warehouse_forklift_routing': Value('int64'), 'wireless_electricity_layout': Value('int64'), 'ordinal_notation_well_foundedness': Value('int64'), 'anchorhead_text_adventure': Value('int64'), 'dcss_dungeon_ai': Value('int64'), 'openttd_transport_ai': Value('int64'), 'tryst_text_adventure': Value('int64'), 'wesnoth_tactical_ai': Value('int64'), 'vibrating_path_graph_coloring': Value('int64'), 'cta_risk_budget_optimization': Value('int64'), 'k12_math_recommendation': Value('int64'), 'portfolio_risk_calibration': Value('int64'), 'combinatorial_games_formalization': Value('int64'), 'new_foundations_consistency': Value('int64'), 'pfr_formalization': Value('int64'), 'sphere_eversion_formalization': Value('int64'), 'trinity_text_adventure': Value('int64'), 'nethack_dungeon_agent': Value('int64'), 'flt_regular_formalization': Value('int64'), 'openrct2_theme_park_ai': Value('int64')}, 'shortened_text_fields': Value('int64'), 'trajectories_unchanged': Value('bool'), 'zip_count': Value('int64'), 'zip_bytes': Value('int64'), 'max_allowed_zip_count': 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 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
              task_count: int64
              tasks_4h: int64
              tasks_8h: int64
              model_calls: int64
              codex_sessions: int64
              model: string
              all_tasks_exit_zero: bool
              score_entries_per_task: struct<arc_compiler_runtime: int64, exchange_core_throughput: int64, integer_compression_codec: int6 (... 1521 chars omitted)
                child 0, arc_compiler_runtime: int64
                child 1, exchange_core_throughput: int64
                child 2, integer_compression_codec: int64
                child 3, vliw_kernel_optimization: int64
                child 4, apple_incremental_game: int64
                child 5, grid_turing_robot: int64
                child 6, molecular_self_assembly: int64
                child 7, order_addition_permutation_optimization: int64
                child 8, treant_forest: int64
                child 9, bipedalwalker_locomotion_rl: int64
                child 10, dabic_gravity_inversion: int64
                child 11, ffmpeg_swscale_reimplementation: int64
                child 12, schemathesis_config_modernization: int64
                child 13, schemathesis_reporting_observability: int64
                child 14, borden_source_inversion: int64
                child 15, juliet_vulnerability_analyzer: int64
                child 16, schemathesis_datagen_pipeline: int64
                child 17, ad_placement_optimization: int64
                child 18, jagua_nesting_optimization: int64
                child 19, git_rewrite_in_zig: int64
                child 20, rust_multicrate_reconstruction: int64
                child 21, equivalence_class_divide_and_conquer: int64
                child 22, graph_node_classification: int64
                child 23, ann_vector_search_qps: int64
                child 24, tree_block_partitioning: int64
                child 25, triangulation_coloring_optimization: int64
                child 26, vehicle_routing_time_windows: int64
                child 27, warehouse_forklift_routing: int64
                child 28, wireless_electricity_layout: int64
                child 29, ordinal_notation_well_foundedness: int64
                child 30, anchorhead_text_adventure: int64
                child 31, dcss_dungeon_ai: int64
                child 32, openttd_transport_ai: int64
                child 33, tryst_text_adventure: int64
                child 34, wesnoth_tactical_ai: int64
                child 35, vibrating_path_graph_coloring: int64
                child 36, cta_risk_budget_optimization: int64
                child 37, k12_math_recommendation: int64
                child 38, portfolio_risk_calibration: int64
                child 39, combinatorial_games_formalization: int64
                child 40, new_foundations_consistency: int64
                child 41, pfr_formalization: int64
                child 42, sphere_eversion_formalization: int64
                child 43, trinity_text_adventure: int64
                child 44, nethack_dungeon_agent: int64
                child 45, flt_regular_formalization: int64
                child 46, openrct2_theme_park_ai: int64
              zip_bytes: int64
              shortened_text_fields: int64
              zip_count: int64
              tasks: int64
              score_entries: int64
              trajectories_unchanged: bool
              max_allowed_zip_count: int64
              to
              {'tasks': Value('int64'), 'model_calls': Value('int64'), 'tasks_4h': Value('int64'), 'tasks_8h': Value('int64'), 'score_entries': Value('int64'), 'score_entries_per_task': {'arc_compiler_runtime': Value('int64'), 'exchange_core_throughput': Value('int64'), 'integer_compression_codec': Value('int64'), 'vliw_kernel_optimization': Value('int64'), 'apple_incremental_game': Value('int64'), 'grid_turing_robot': Value('int64'), 'molecular_self_assembly': Value('int64'), 'order_addition_permutation_optimization': Value('int64'), 'treant_forest': Value('int64'), 'bipedalwalker_locomotion_rl': Value('int64'), 'dabic_gravity_inversion': Value('int64'), 'ffmpeg_swscale_reimplementation': Value('int64'), 'schemathesis_config_modernization': Value('int64'), 'schemathesis_reporting_observability': Value('int64'), 'borden_source_inversion': Value('int64'), 'juliet_vulnerability_analyzer': Value('int64'), 'schemathesis_datagen_pipeline': Value('int64'), 'ad_placement_optimization': Value('int64'), 'jagua_nesting_optimization': Value('int64'), 'git_rewrite_in_zig': Value('int64'), 'rust_multicrate_reconstruction': Value('int64'), 'equivalence_class_divide_and_conquer': Value('int64'), 'graph_node_classification': Value('int64'), 'ann_vector_search_qps': Value('int64'), 'tree_block_partitioning': Value('int64'), 'triangulation_coloring_optimization': Value('int64'), 'vehicle_routing_time_windows': Value('int64'), 'warehouse_forklift_routing': Value('int64'), 'wireless_electricity_layout': Value('int64'), 'ordinal_notation_well_foundedness': Value('int64'), 'anchorhead_text_adventure': Value('int64'), 'dcss_dungeon_ai': Value('int64'), 'openttd_transport_ai': Value('int64'), 'tryst_text_adventure': Value('int64'), 'wesnoth_tactical_ai': Value('int64'), 'vibrating_path_graph_coloring': Value('int64'), 'cta_risk_budget_optimization': Value('int64'), 'k12_math_recommendation': Value('int64'), 'portfolio_risk_calibration': Value('int64'), 'combinatorial_games_formalization': Value('int64'), 'new_foundations_consistency': Value('int64'), 'pfr_formalization': Value('int64'), 'sphere_eversion_formalization': Value('int64'), 'trinity_text_adventure': Value('int64'), 'nethack_dungeon_agent': Value('int64'), 'flt_regular_formalization': Value('int64'), 'openrct2_theme_park_ai': Value('int64')}, 'shortened_text_fields': Value('int64'), 'trajectories_unchanged': Value('bool'), 'zip_count': Value('int64'), 'zip_bytes': Value('int64'), 'max_allowed_zip_count': 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.

EdgeBench rollouts: Opus 5.5, Kimi-K3, Qwen3.8 Flash Next and Qwen3.5

Claude Code scaffold, model claude-opus-5-5, effort xhigh, 4 hours of agent time per task. Run ID opus-16x4h-1009-0036, waves w1-w4; one directory per task with the transcript (agent_output.txt), submissions, final archive and results (final_result.json). The 16 tasks are available as 6 independent ZIP files, each strictly below 512 MB (512,000,000 bytes). See the Opus ZIP download and task index. The old oversized .tar.zst has been replaced. Each ZIP opens independently; these are not multipart volumes.

Notes: lean_analysis_proofs was paused after ~11 min and resumed later (resume_count 5); the 4 h budget counts only time the agent actually ran. flt_regular_formalization ended itself after ~2 h.

Kimi-K3 rollouts (Codex scaffold), kimi-k3/

kimi-k3/kimi-full50-rollouts-light.tar.gz (134 MB, tar -xzf; SHA-256 eb779d62a7d230b35b6cffe2d1f880f4589bb483ca3545ac48da754a4699f8ec): three runs of Kimi-K3 on 50 EdgeBench tasks, Codex scaffold, 4 h of agent time per task:

  • full50-20260908-213644: reasoning effort not forwarded (server default);
  • full50-max-20260909-230641: effort max (includes retries);
  • full50-max-cont8h-20260911-012000: continuation of the max run to 8 h.

Layout: <run>/harness/runs/<run>-<task>/<task>/ with Codex's plain-text log (agent_output.txt), run_history.json, final_result.json and submissions/<round>/report.json. Light copy: the per-task submission.tar.gz and final_archive.tar.gz and the full model exchanges are left out (the full set is about 53 GB). These runs predate a fix in our Responses bridge: earlier turns' reasoning was dropped from the history, so most turns have no reasoning text.

Qwen3.8 Flash Next rollouts (Codex scaffold), qwen3.8-flash-next/

All 47 tasks (24 from the original batch plus 23 additional tasks) are now available as 2 independent ZIP files, each strictly below 512 MB (512,000,000 bytes). See the Qwen ZIP download and task index. The old oversized 47-task .tar.gz and first-batch archive have been replaced. Each ZIP can be opened independently. Extract all ZIPs into the same directory to recover the complete export-full47/ layout; no volume concatenation is needed. Tasks may span ZIPs to minimize the number of packages. Every individual file is intact. The task and file indices show where to find each item.

Official Qwen3.8-Flash-Next-FP8, vLLM 0.31.0, one GPU per instance, MTP speculative decoding, Codex scaffold, effort xhigh, native 262144-token context. Run qwen38fn-full48-12h-20261015, 12-hour budget per task.

Contents across the ZIPs:

  • trajectories/<task>.trajectory.jsonl.gz: 47 unique reconstructed trajectories, with 130,361 model calls in total. Recorded reasoning, messages, tool calls, tool outputs and context transitions are retained. All model-call IDs in the original 24-task batch are retained.
  • harness/<task>/: prompts, agent logs, evaluation histories, final results and final code archives where produced. Text-adventure tasks use game_history.json.
  • codex-sessions/<task>/: 51 original Codex session JSONL files for 47 tasks, including resumed sessions.
  • task_index.json and final_results_47.json: all 47 tasks and their final results.
  • scores_4h_8h_12h.json: the existing checkpoint-score export for only the first 24 tasks; no batch2 checkpoint scores are inferred.
  • Serving/queue metadata and incident notes.

Known runtime access tokens are replaced with equal-length x characters in the publication copy. The original local data is unchanged. Repeated full-history HTTP captures and container filesystem-layer archives are omitted from the compact publication; reconstructed trajectories and original session logs are included.

The 12-hour budget does not imply 12 hours of productive work. Recorded hangs / PID exhaustion affected order_addition_permutation_optimization, nethack_dungeon_agent and dcss_dungeon_ai; see incident notes. lean_analysis_proofs was dropped and is not among the 47 tasks; smt_solver, carleson_formalization and college_english_exam_bank were also excluded. The agent of schemathesis_config_modernization called the model API itself with curl a few times.

Qwen3.5-122B-A10B rollouts (Codex scaffold), qwen3.5-122b-a10b/

For ChatGPT analysis: use the single compact ZIP. It contains all 47 full trajectories and concise scores for every submission/game round, below 512 MB. The 24-ZIP full archive below is optional for analysis.

47 tasks: 24 at 4 hours and 23 at 8 hours. The first 24 were not continued to 8 hours. All 47 harness jobs exited successfully. This is a runtime-completion statement, not a task-success score. The exact per-task budgets are in task_index.json.

Available as 24 independent ZIP files, each strictly below 512,000,000 bytes. This reaches the minimum package count for these compressed files. See the Qwen3.5 ZIP downloads and task locations. Extract all ZIPs into the same directory to restore export-qwen35-full47/. Each individual file remains intact; tasks may span ZIPs.

Model Qwen3.5-122B-A10B (BF16), vLLM, 8 single-GPU instances, Codex scaffold, thinking enabled, native 262144-token context, no fixed reasoning-effort override. This is separate from the Qwen3.8 Flash Next data above.

Contents: 88,109 model calls, 47 compact reconstructed trajectories, 47 original Codex session JSONL files, prompts and agent logs, submission/judge details, final workspace archives, final results, and both batches' run metadata. Reasoning, tool calls, tool outputs and context transitions are retained. Known runtime credentials are redacted from the publication copy. Repeated full-history HTTP captures and container filesystem-layer images are omitted from this compact publication; raw HTTP captures remain in the workstation backup.

Excluded: lean_analysis_proofs, smt_solver, carleson_formalization, college_english_exam_bank.

Downloads last month
1,310