Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
case_id: string
construction: struct<candidate_id: string, edge_semantics: string, whole_graph: string>
  child 0, candidate_id: string
  child 1, edge_semantics: string
  child 2, whole_graph: string
edges: list<item: struct<evidence: string, from: string, to: string>>
  child 0, item: struct<evidence: string, from: string, to: string>
      child 0, evidence: string
      child 1, from: string
      child 2, to: string
math_case_ref: string
nodes: list<item: struct<gold_hidden_from_agents: bool, id: string, module: string, source_block_sha256: st (... 106 chars omitted)
  child 0, item: struct<gold_hidden_from_agents: bool, id: string, module: string, source_block_sha256: string, sourc (... 94 chars omitted)
      child 0, gold_hidden_from_agents: bool
      child 1, id: string
      child 2, module: string
      child 3, source_block_sha256: string
      child 4, source_file: string
      child 5, source_line: int64
      child 6, stage: int64
      child 7, statement: string
      child 8, statement_sha256: string
proof_contract: struct<>
source: struct<commit: string, lean_toolchain: string, project: string, project_id: string>
  child 0, commit: string
  child 1, lean_toolchain: string
  child 2, project: string
  child 3, project_id: string
topology: struct<depth: int64, edge_count: int64, join_nodes: list<item: string>, kind: string, node_count: in (... 86 chars omitted)
  child 0, depth: int64
  child 1, edge_count: int64
  child 2, join_nodes: list<item: string>
      child 0, item: string
  child 3, kind: string
  child 4, node_count: int64
  child 5, overlap_nodes: list<item: string>
      child 0, item: string
  child 6, reuse_nodes: list<item: string>
      child 0, item: string
  child 7, width: int64
instance_id: string
control_instance_id: string
family_id: string
variant: string
scenario_id: string
schema_version: string
pollution_plan_ref: string
pair_id: string
to
{'control_instance_id': Value('string'), 'family_id': Value('string'), 'instance_id': Value('string'), 'pair_id': Value('string'), 'pollution_plan_ref': Value('string'), 'scenario_id': Value('string'), 'schema_version': Value('string'), 'variant': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                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
              case_id: string
              construction: struct<candidate_id: string, edge_semantics: string, whole_graph: string>
                child 0, candidate_id: string
                child 1, edge_semantics: string
                child 2, whole_graph: string
              edges: list<item: struct<evidence: string, from: string, to: string>>
                child 0, item: struct<evidence: string, from: string, to: string>
                    child 0, evidence: string
                    child 1, from: string
                    child 2, to: string
              math_case_ref: string
              nodes: list<item: struct<gold_hidden_from_agents: bool, id: string, module: string, source_block_sha256: st (... 106 chars omitted)
                child 0, item: struct<gold_hidden_from_agents: bool, id: string, module: string, source_block_sha256: string, sourc (... 94 chars omitted)
                    child 0, gold_hidden_from_agents: bool
                    child 1, id: string
                    child 2, module: string
                    child 3, source_block_sha256: string
                    child 4, source_file: string
                    child 5, source_line: int64
                    child 6, stage: int64
                    child 7, statement: string
                    child 8, statement_sha256: string
              proof_contract: struct<>
              source: struct<commit: string, lean_toolchain: string, project: string, project_id: string>
                child 0, commit: string
                child 1, lean_toolchain: string
                child 2, project: string
                child 3, project_id: string
              topology: struct<depth: int64, edge_count: int64, join_nodes: list<item: string>, kind: string, node_count: in (... 86 chars omitted)
                child 0, depth: int64
                child 1, edge_count: int64
                child 2, join_nodes: list<item: string>
                    child 0, item: string
                child 3, kind: string
                child 4, node_count: int64
                child 5, overlap_nodes: list<item: string>
                    child 0, item: string
                child 6, reuse_nodes: list<item: string>
                    child 0, item: string
                child 7, width: int64
              instance_id: string
              control_instance_id: string
              family_id: string
              variant: string
              scenario_id: string
              schema_version: string
              pollution_plan_ref: string
              pair_id: string
              to
              {'control_instance_id': Value('string'), 'family_id': Value('string'), 'instance_id': Value('string'), 'pair_id': Value('string'), 'pollution_plan_ref': Value('string'), 'scenario_id': Value('string'), 'schema_version': Value('string'), 'variant': Value('string')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1879, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

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.

control_instance_id
string
family_id
string
instance_id
string
pair_id
string
pollution_plan_ref
string
scenario_id
string
schema_version
string
variant
string
null
v15_family_0001_b9f32b72e8fd
v15_family_0001_b9f32b72e8fd::control
v24_pair_1963a85afa2b
null
v15_scenario_7ca77aa1c77f
comembench-instance/v17
control
null
v15_family_0002_1a53dcfbb148
v15_family_0002_1a53dcfbb148::control
v24_pair_b9dafa7f33ef
null
v15_scenario_ab135718378f
comembench-instance/v17
control
null
v15_family_0003_0b3c745bd39a
v15_family_0003_0b3c745bd39a::control
v24_pair_2ebb0d6af6d6
null
v15_scenario_2dfe376f24b5
comembench-instance/v17
control
null
v15_family_0004_ba2cbaee8dc8
v15_family_0004_ba2cbaee8dc8::control
v24_pair_5c6eae53cb57
null
v15_scenario_02518fb53c7a
comembench-instance/v17
control
null
v15_family_0005_f30c9780119b
v15_family_0005_f30c9780119b::control
v24_pair_b27cd962f933
null
v15_scenario_435ac4148a8f
comembench-instance/v17
control
null
v15_family_0006_c1c2de9befcb
v15_family_0006_c1c2de9befcb::control
v24_pair_3e721eef6e3b
null
v15_scenario_ff5d2ddd4ef1
comembench-instance/v17
control
null
v15_family_0007_91a216b96713
v15_family_0007_91a216b96713::control
v24_pair_4a1d12c56835
null
v15_scenario_ce34270655b8
comembench-instance/v17
control
null
v15_family_0008_45dd6334445b
v15_family_0008_45dd6334445b::control
v24_pair_dafea55eba6c
null
v15_scenario_303515805c9b
comembench-instance/v17
control
null
v15_family_0009_de87d96c105d
v15_family_0009_de87d96c105d::control
v24_pair_1d6d2e9cc502
null
v15_scenario_fade350297cd
comembench-instance/v17
control
null
v15_family_0010_76b471c05a36
v15_family_0010_76b471c05a36::control
v24_pair_f7a9552a8461
null
v15_scenario_46cc01cd9496
comembench-instance/v17
control
null
v15_family_0011_f10b39a45eee
v15_family_0011_f10b39a45eee::control
v24_pair_0c99a1feb99e
null
v15_scenario_18bb11b9aba5
comembench-instance/v17
control
null
v15_family_0012_3b67a55f0620
v15_family_0012_3b67a55f0620::control
v24_pair_d3599a509901
null
v15_scenario_4fd6d29cf9d2
comembench-instance/v17
control
null
v15_family_0013_ee55ec32355e
v15_family_0013_ee55ec32355e::control
v24_pair_7de2fa4f87c5
null
v15_scenario_352a802b0e39
comembench-instance/v17
control
null
v15_family_0014_82492815fe64
v15_family_0014_82492815fe64::control
v24_pair_cdc25d9ee121
null
v15_scenario_a61ee96b004e
comembench-instance/v17
control
null
v15_family_0015_8fc04dfddff9
v15_family_0015_8fc04dfddff9::control
v24_pair_b453b875b4f1
null
v15_scenario_3ad8aea8c7f4
comembench-instance/v17
control
null
v15_family_0016_4ac1a5cf5191
v15_family_0016_4ac1a5cf5191::control
v24_pair_848fc2f99dab
null
v15_scenario_b548aa7615ba
comembench-instance/v17
control
null
v15_family_0017_fab8427d066b
v15_family_0017_fab8427d066b::control
v24_pair_c6c4bddc6668
null
v15_scenario_ba5fdcc61f91
comembench-instance/v17
control
null
v15_family_0018_57541cbaddf2
v15_family_0018_57541cbaddf2::control
v24_pair_80012c5b3849
null
v15_scenario_14bac8816673
comembench-instance/v17
control
null
v15_family_0019_b62475916eca
v15_family_0019_b62475916eca::control
v24_pair_bfd0b5ee16ef
null
v15_scenario_1d82dcf4b280
comembench-instance/v17
control
null
v15_family_0020_37e20e8c9cc2
v15_family_0020_37e20e8c9cc2::control
v24_pair_c96229e7c11f
null
v15_scenario_4d432ba4d7b7
comembench-instance/v17
control
null
v15_family_0021_8c13d2c71488
v15_family_0021_8c13d2c71488::control
v24_pair_92105717840c
null
v15_scenario_40fd9e3f4505
comembench-instance/v17
control
null
v15_family_0022_bfd24cb7f0cf
v15_family_0022_bfd24cb7f0cf::control
v24_pair_d14a1ffaeb98
null
v15_scenario_cfec238ab628
comembench-instance/v17
control
null
v15_family_0023_1bbefff7c7b7
v15_family_0023_1bbefff7c7b7::control
v24_pair_aa687e6cdfe2
null
v15_scenario_70d8eca34adc
comembench-instance/v17
control
null
v15_family_0024_06d4920d28ea
v15_family_0024_06d4920d28ea::control
v24_pair_fd57c1f39721
null
v15_scenario_5858ddf84ffc
comembench-instance/v17
control
null
v15_family_0025_bbb3dd9c566e
v15_family_0025_bbb3dd9c566e::control
v24_pair_4734bc3e9075
null
v15_scenario_7dabba2c806e
comembench-instance/v17
control
null
v15_family_0026_a8e90faf7905
v15_family_0026_a8e90faf7905::control
v24_pair_554a44d54ba5
null
v15_scenario_12a14d7fed49
comembench-instance/v17
control
null
v15_family_0027_2d10ae8c60ca
v15_family_0027_2d10ae8c60ca::control
v24_pair_17757078d3cb
null
v15_scenario_ee40330e4919
comembench-instance/v17
control
null
v15_family_0028_5899ddc138f2
v15_family_0028_5899ddc138f2::control
v24_pair_4638c0495a23
null
v15_scenario_e622b3b156f9
comembench-instance/v17
control
null
v15_family_0029_a589445a5ec4
v15_family_0029_a589445a5ec4::control
v24_pair_fce03e059b9b
null
v15_scenario_e0bef5b3ef9c
comembench-instance/v17
control
null
v15_family_0030_195cd2233b99
v15_family_0030_195cd2233b99::control
v24_pair_2bbddc467905
null
v15_scenario_252ddc56b7c7
comembench-instance/v17
control
null
v15_family_0031_2e00a9d15ff6
v15_family_0031_2e00a9d15ff6::control
v24_pair_0db5d74ed8b1
null
v15_scenario_70a89a0c7c16
comembench-instance/v17
control
null
v15_family_0032_8b3bf78e3e06
v15_family_0032_8b3bf78e3e06::control
v24_pair_db763f7e38d0
null
v15_scenario_160c8b49184f
comembench-instance/v17
control
null
v15_family_0033_058a9d3591e4
v15_family_0033_058a9d3591e4::control
v24_pair_a8840edd376b
null
v15_scenario_6f5dcd4039dd
comembench-instance/v17
control
null
v15_family_0034_90a978eaead5
v15_family_0034_90a978eaead5::control
v24_pair_000c61a4e605
null
v15_scenario_cd6d4bf87008
comembench-instance/v17
control
null
v15_family_0035_951ccdc14494
v15_family_0035_951ccdc14494::control
v24_pair_5252de55fd4f
null
v15_scenario_cdbb61fc75ed
comembench-instance/v17
control
null
v15_family_0036_75bd149dd987
v15_family_0036_75bd149dd987::control
v24_pair_2c15488e23c6
null
v15_scenario_29850738a60a
comembench-instance/v17
control
null
v15_family_0037_bbf1ad486f54
v15_family_0037_bbf1ad486f54::control
v24_pair_94a4034fcfe0
null
v15_scenario_33aafa08caaa
comembench-instance/v17
control
null
v15_family_0038_268e41ee3aed
v15_family_0038_268e41ee3aed::control
v24_pair_705d57386d79
null
v15_scenario_c902bad40e86
comembench-instance/v17
control
null
v15_family_0039_49f787f0f704
v15_family_0039_49f787f0f704::control
v24_pair_3d169ddd1058
null
v15_scenario_d32bb16227ad
comembench-instance/v17
control
null
v15_family_0040_4c9f7127fbe2
v15_family_0040_4c9f7127fbe2::control
v24_pair_25f5a7b083c5
null
v15_scenario_b9b6696e342d
comembench-instance/v17
control
null
v15_family_0041_505e453742a4
v15_family_0041_505e453742a4::control
v24_pair_e64864bab1e8
null
v15_scenario_0cfc153c0fd7
comembench-instance/v17
control
null
v15_family_0042_8c45031ee1ec
v15_family_0042_8c45031ee1ec::control
v24_pair_74d3f8e9a7c3
null
v15_scenario_68de4959fa08
comembench-instance/v17
control
null
v15_family_0043_6964fcd88437
v15_family_0043_6964fcd88437::control
v24_pair_05a51f096cf4
null
v15_scenario_e271f81ea8f5
comembench-instance/v17
control
null
v15_family_0044_34d265afd8ff
v15_family_0044_34d265afd8ff::control
v24_pair_20c55ccf428a
null
v15_scenario_35d4b79f152c
comembench-instance/v17
control
null
v15_family_0045_000a1689f1a6
v15_family_0045_000a1689f1a6::control
v24_pair_fabf3bef3fe3
null
v15_scenario_8873aea3f237
comembench-instance/v17
control
null
v15_family_0046_9aebaecb4293
v15_family_0046_9aebaecb4293::control
v24_pair_b7f9dd8bad68
null
v15_scenario_58c7fe41b81d
comembench-instance/v17
control
null
v15_family_0047_d846ce2c9c89
v15_family_0047_d846ce2c9c89::control
v24_pair_5d49e1fcda07
null
v15_scenario_3f729af75b6a
comembench-instance/v17
control
null
v15_family_0048_5fcfc655726f
v15_family_0048_5fcfc655726f::control
v24_pair_558d115b7ee4
null
v15_scenario_294d31c1f46c
comembench-instance/v17
control
null
v15_family_0049_bed5006a73db
v15_family_0049_bed5006a73db::control
v24_pair_3867f6013dd1
null
v15_scenario_d96ed058df8d
comembench-instance/v17
control
null
v15_family_0050_cced0e6690e5
v15_family_0050_cced0e6690e5::control
v24_pair_f7af93b23fc1
null
v15_scenario_d140dbc5a537
comembench-instance/v17
control
null
v15_family_0051_417ca099bf58
v15_family_0051_417ca099bf58::control
v24_pair_60fabdbb36d8
null
v15_scenario_a02e479a29b8
comembench-instance/v17
control
null
v15_family_0052_a82adf411e9f
v15_family_0052_a82adf411e9f::control
v24_pair_af20272ceb88
null
v15_scenario_b36c782330e9
comembench-instance/v17
control
null
v15_family_0053_6a49be232dbe
v15_family_0053_6a49be232dbe::control
v24_pair_3694caeeec3e
null
v15_scenario_6b072bcf4347
comembench-instance/v17
control
null
v15_family_0054_6f57c1c56db7
v15_family_0054_6f57c1c56db7::control
v24_pair_84779a6257e9
null
v15_scenario_545e82b18e36
comembench-instance/v17
control
null
v15_family_0055_a8422775aaad
v15_family_0055_a8422775aaad::control
v24_pair_19218ad94edf
null
v15_scenario_132c4c9a0110
comembench-instance/v17
control
null
v15_family_0056_e204dc664fad
v15_family_0056_e204dc664fad::control
v24_pair_a383284ad7d6
null
v15_scenario_7496cdc02e46
comembench-instance/v17
control
null
v15_family_0057_d966ce53d267
v15_family_0057_d966ce53d267::control
v24_pair_961390bc2d19
null
v15_scenario_a0b2268f6429
comembench-instance/v17
control
null
v15_family_0058_12a91172c340
v15_family_0058_12a91172c340::control
v24_pair_9cda177d14dd
null
v15_scenario_6ad61fec3bce
comembench-instance/v17
control
null
v15_family_0059_43b2fa5c33a4
v15_family_0059_43b2fa5c33a4::control
v24_pair_cadaa35274ab
null
v15_scenario_15a5084eeff5
comembench-instance/v17
control
null
v15_family_0060_4d10a8eea0da
v15_family_0060_4d10a8eea0da::control
v24_pair_cb3dad2e876e
null
v15_scenario_bd4621ec21a0
comembench-instance/v17
control
null
v15_family_0061_de7fac75a9d2
v15_family_0061_de7fac75a9d2::control
v24_pair_78305100aeae
null
v15_scenario_00745f5f3e57
comembench-instance/v17
control
null
v15_family_0062_ea63efcfc385
v15_family_0062_ea63efcfc385::control
v24_pair_a3c3d23396f2
null
v15_scenario_cc1b353b0fda
comembench-instance/v17
control
null
v15_family_0063_066a1762d8a2
v15_family_0063_066a1762d8a2::control
v24_pair_187f9b2369fe
null
v15_scenario_e52c20765c8f
comembench-instance/v17
control
null
v15_family_0064_40edfad58e89
v15_family_0064_40edfad58e89::control
v24_pair_d2a4ca40f848
null
v15_scenario_7eb7d061f5b4
comembench-instance/v17
control
null
v15_family_0065_60bb20520916
v15_family_0065_60bb20520916::control
v24_pair_99a9f067af6c
null
v15_scenario_01217ec088a4
comembench-instance/v17
control
null
v15_family_0066_8c10e1ab5cc4
v15_family_0066_8c10e1ab5cc4::control
v24_pair_e3d88ab0ce47
null
v15_scenario_979ae1d44099
comembench-instance/v17
control
null
v15_family_0067_5a82fa624694
v15_family_0067_5a82fa624694::control
v24_pair_9f671799874b
null
v15_scenario_d3bb1e91bf87
comembench-instance/v17
control
null
v15_family_0068_d337d5b0744f
v15_family_0068_d337d5b0744f::control
v24_pair_c302ce6aa121
null
v15_scenario_5a1928cd6a3b
comembench-instance/v17
control
null
v15_family_0069_1d4724ff2154
v15_family_0069_1d4724ff2154::control
v24_pair_215e1ed8d5f6
null
v15_scenario_de8393a94965
comembench-instance/v17
control
null
v15_family_0070_e7e28b3e3a75
v15_family_0070_e7e28b3e3a75::control
v24_pair_bfe939e1bdef
null
v15_scenario_ae73d2cd958a
comembench-instance/v17
control
null
v15_family_0071_543136082c1a
v15_family_0071_543136082c1a::control
v24_pair_3d9e3a3d51e9
null
v15_scenario_dbec4622f64a
comembench-instance/v17
control
null
v15_family_0072_763351973dbf
v15_family_0072_763351973dbf::control
v24_pair_7047cdb04233
null
v15_scenario_a5b516919a33
comembench-instance/v17
control
null
v15_family_0073_1e248b8fd711
v15_family_0073_1e248b8fd711::control
v24_pair_5cd6987960b2
null
v15_scenario_d7f04b1ae828
comembench-instance/v17
control
null
v15_family_0074_c11c33926c2b
v15_family_0074_c11c33926c2b::control
v24_pair_1e7eddf84cc9
null
v15_scenario_83438b4ad16b
comembench-instance/v17
control
null
v15_family_0075_17497e021baa
v15_family_0075_17497e021baa::control
v24_pair_9576d075f90b
null
v15_scenario_035c28122fdc
comembench-instance/v17
control
null
v15_family_0076_28374ef5e8fc
v15_family_0076_28374ef5e8fc::control
v24_pair_592792fb854e
null
v15_scenario_bf7ad3a4cec2
comembench-instance/v17
control
null
v15_family_0077_a6fb18f9b3f6
v15_family_0077_a6fb18f9b3f6::control
v24_pair_219adb109326
null
v15_scenario_21756ea5451d
comembench-instance/v17
control
null
v15_family_0078_156b76928343
v15_family_0078_156b76928343::control
v24_pair_9323a7ade39f
null
v15_scenario_447cea1ff323
comembench-instance/v17
control
null
v15_family_0079_c70e5d3b66e9
v15_family_0079_c70e5d3b66e9::control
v24_pair_6966ae26afee
null
v15_scenario_5244cd5065a4
comembench-instance/v17
control
null
v15_family_0080_cec66303c0b9
v15_family_0080_cec66303c0b9::control
v24_pair_adcfbafa3a3d
null
v15_scenario_667c5cc9b707
comembench-instance/v17
control
null
v15_family_0081_09736c633fcd
v15_family_0081_09736c633fcd::control
v24_pair_2a5e70e1d0bb
null
v15_scenario_0f9a8b518cb3
comembench-instance/v17
control
null
v15_family_0082_69b6024d3867
v15_family_0082_69b6024d3867::control
v24_pair_562ae3bbf8b0
null
v15_scenario_e1f7b79b26ec
comembench-instance/v17
control
null
v15_family_0083_7e674be31a46
v15_family_0083_7e674be31a46::control
v24_pair_f185aff8693f
null
v15_scenario_2f744b62ed91
comembench-instance/v17
control
null
v15_family_0084_fe19e8c20cea
v15_family_0084_fe19e8c20cea::control
v24_pair_6b3cd73c8dea
null
v15_scenario_f05c3ad63caa
comembench-instance/v17
control
null
v15_family_0085_b3cae95c4901
v15_family_0085_b3cae95c4901::control
v24_pair_dc051516b946
null
v15_scenario_003bb58622d1
comembench-instance/v17
control
null
v15_family_0086_55e2bbcab08b
v15_family_0086_55e2bbcab08b::control
v24_pair_8f4fc9c802fa
null
v15_scenario_8d9ac11f2e78
comembench-instance/v17
control
null
v15_family_0087_bef370ffa855
v15_family_0087_bef370ffa855::control
v24_pair_8d4e443fd52a
null
v15_scenario_82f53cdb573c
comembench-instance/v17
control
null
v15_family_0088_aa09db125062
v15_family_0088_aa09db125062::control
v24_pair_3503287ecc40
null
v15_scenario_0cf775d488da
comembench-instance/v17
control
null
v15_family_0089_f9b43078bdb8
v15_family_0089_f9b43078bdb8::control
v24_pair_3b829d8b261d
null
v15_scenario_d25e0a0ed1ad
comembench-instance/v17
control
null
v15_family_0090_9338d6928225
v15_family_0090_9338d6928225::control
v24_pair_c8bd49cdb774
null
v15_scenario_364d13e428cd
comembench-instance/v17
control
null
v15_family_0091_b7b1912cf711
v15_family_0091_b7b1912cf711::control
v24_pair_1930b9a845ed
null
v15_scenario_fe418c890ea1
comembench-instance/v17
control
null
v15_family_0092_0ed0f7bf2fa0
v15_family_0092_0ed0f7bf2fa0::control
v24_pair_7721aad6e0d8
null
v15_scenario_2104a07d7fda
comembench-instance/v17
control
null
v15_family_0093_97fc9ccc5861
v15_family_0093_97fc9ccc5861::control
v24_pair_bba623fb4e67
null
v15_scenario_54a806de7db5
comembench-instance/v17
control
null
v15_family_0094_899ccd1982f6
v15_family_0094_899ccd1982f6::control
v24_pair_4d0600847c59
null
v15_scenario_a2d906d90e01
comembench-instance/v17
control
null
v15_family_0095_a7cbcd20f7e0
v15_family_0095_a7cbcd20f7e0::control
v24_pair_7618d9b9fdcc
null
v15_scenario_7b68fec8bd1e
comembench-instance/v17
control
null
v15_family_0096_81bcc8355b35
v15_family_0096_81bcc8355b35::control
v24_pair_5ab7b7c8035f
null
v15_scenario_47d6a7fb3f18
comembench-instance/v17
control
null
v15_family_0097_e1681e3d573f
v15_family_0097_e1681e3d573f::control
v24_pair_b0f4b090f1dd
null
v15_scenario_88b5021a6923
comembench-instance/v17
control
null
v15_family_0098_87c72570ac78
v15_family_0098_87c72570ac78::control
v24_pair_39e555fa9c6a
null
v15_scenario_7805c76b5c84
comembench-instance/v17
control
null
v15_family_0099_018c65119712
v15_family_0099_018c65119712::control
v24_pair_84b9e646e136
null
v15_scenario_34191a38ed43
comembench-instance/v17
control
null
v15_family_0100_e9c604baf2fd
v15_family_0100_e9c604baf2fd::control
v24_pair_27b62c837b3a
null
v15_scenario_06ae79ead79f
comembench-instance/v17
control
End of preview.

CoMemBench: Benchmarking Collaborative Memory Boundaries across Multi-Agent Workflow Topologies

CoMemBench is a benchmark for collaborative memory in multi-agent systems. Each problem is a real multi-agent workflow β€” a directed graph of specialized worker nodes with typed handoffs β€” and the benchmark asks whether a system can run that workflow while using only the memory it should use: no stale artifacts, no cross-wired branches, no foreign completions.

The stress is a matched pollution variant. For every problem there is a twin that differs in exactly one architecture-visible way: one additional routing candidate is offered to the workflow, and that candidate is invalid. A system that reads memory indiscriminately will take the bait and fail; a system with correct memory boundaries will reject it and finish.

This repository is the anonymised artifact accompanying a manuscript under double-blind review. All author and affiliation information has been removed; it will be published under the authors' names upon acceptance.

At a glance

Complete problems 800 (4 domains Γ— 200)
Matched polluted instances 800 (one per problem)
Total instances 1600 (800 control + 800 polluted)
Workflow nodes per problem 6–19 (median 11)
Node-level memory contexts 8975
Held-out split none β€” all 800 are complete problems
Version v1

Domains

Domain Count Node style
stateful_tool_use 200 Stateful tool chains β€” terminal, documents, spreadsheets, forms
repository_code_change 200 Multi-node code-change workflows over real repositories
offline_information_retrieval 200 Multi-hop retrieval pipelines over fixed corpora
mathematical_and_structured_reasoning 200 Proof-dependent and structured-reasoning workflows

What a problem looks like

A workflow family is a layered DAG of worker nodes. Each node carries:

  • a role and a task (what this worker must produce),
  • a local context (the memory scope this worker is allowed to read),
  • an output contract (artifact semantic + runtime contract),
  • task edges describing the handoffs (artifact reuse, verification, dependency evidence).

Plus a topology motif set (chain, split_join, reuse, overlap, bounded_revision_loop β€” most families are composite) and a revision contract that bounds how many times the workflow may revisit work.

The agent's job is to execute the workflow: produce artifacts, hand them off along required edges, and satisfy the domain's native evaluator.

Files

data/
  workflow_families.jsonl       800    the problems (goals, nodes, edges, topology, revision contract)
  node_contexts.jsonl          8975    per-node memory scope (environment ref, resource policy, task scope)
  topology_descriptors.jsonl    800    per-family topology statistics (depth, width, fan-out, joins, reuse, overlap)
  math_cases.jsonl              200    Lean dependency cases for the math domain (statements + proof-term DAG)
  retrieval_corpora.jsonl       200    frozen offline corpora for the retrieval domain (one per family)
  control_instances.jsonl       800    control-side instance index
  polluted_instances.jsonl      800    polluted-side instance index (pairs to control via pair_id)
  benchmark_instances.jsonl    1600    both sides merged
schemas/
  workflow_family.schema.json      node_context.schema.json
  benchmark_instance.schema.json   artifact.schema.json
  node_view.schema.json            case_init.schema.json
  memory_proposal.schema.json      handoff_oracle.schema.json
  swe_image_manifest.json          the 50 certified code-domain images (name + sha256 + size)

All files are JSON Lines (one JSON object per line), UTF-8.

data/workflow_families.jsonl

One record per complete problem.

Field Type Meaning
family_id str Stable problem id (e.g. v15_family_0001_b9f32b72e8fd)
scenario_id str Scenario id shared with node contexts
domain str One of the four domains above
global_goal str The overall objective handed to the system
nodes list Worker nodes: node_id, agent_role, task, node_kind, local_context_ref, output_schema, output_semantic, runtime_contract, source_ref, source_unit_id
task_edges list Handoffs: edge_id, from, to, relation_type, handoff_policy, handoff_required, artifact_semantic, dependency_evidence
source_topology / execution_topology dict motifs, primary, signature
revision_contract dict Which edges may be revisited and how many traversals are allowed
execution dict Evaluator mode, revision limit, closed-loop flag
verification_overlay dict Nodes/edges that carry verification duties
construction dict Provenance of how the family was assembled
source_refs list Upstream source records

data/node_contexts.jsonl

One record per workflow node (node_id ↔ context_ref). The local_context field is the memory scope that node is permitted to read: environment_ref, resource_policy, task_specific_scope, source_ref.

data/topology_descriptors.jsonl

One record per family. Structural statistics used for topology-sliced reporting: node_count, edge_count, depth, max_width, fan_out_count, join_count, reuse_count, overlap_count, revision_count, handoff_count, handoff_density.

data/math_cases.jsonl

One record per math-domain family. Nodes are Lean declaration statements (what must be proven) with source location and statement fingerprints; edges are the elaborated proof-term constant dependencies; source pins the project and toolchain. Proof bodies are not included β€” those are the answers.

data/retrieval_corpora.jsonl

One record per retrieval-domain family: corpus_ref, source_ref, and a documents list (docid, text) forming that family's frozen offline corpus. Retrieval problems are answerable only against this corpus.

Instance index files

Thin id-linking tables. control_instances.jsonl and polluted_instances.jsonl share pair_id; polluted_instances.jsonl additionally carries control_instance_id and a pollution_plan_ref (an opaque identifier β€” the plan itself is not part of this release).

Execution environments

CoMemBench runs across four domains with different execution boundaries. Two are self-contained in this repository; two need container images.

Domain Boundary Ships here Needs
mathematical_and_structured_reasoning Host-side frozen Lean declarations math_cases.jsonl leanprover/lean4:v4.7.0 + the pinned HepLean project
offline_information_retrieval Host-side frozen offline corpus retrieval_corpora.jsonl nothing else
repository_code_change One container per case problem definitions the 50 certified images (registry link withheld during review β€” see below)
stateful_tool_use Postgres + private network + MCP processes per task problem definitions postgres:15 (public, Docker Hub) + the Toolathlon-Gym runtime and MCP server set

Code-domain images (50, 333 GB)

The code-change images are the SWE-Verified execution environments (~6–7 GB each), distributed as container images rather than inside this dataset repository β€” the same split SWE-bench uses. All 50 are published as one container package with 50 tags; the tag is the certified image name with the namespace stripped and punctuation normalised (e.g. logicstar/sweb.eval.x86_64.django_1776_django-10097:latest β†’ django-1776-django-10097).

The public registry location is withheld during double-blind review (the URL identifies the authors); it will be listed here upon release. schemas/swe_image_manifest.json pins the exact set: for each image the certified name, its sha256 id, execution_fingerprint, logical size, and slug. When the registry location is public, verify a pulled image against the manifest with docker image inspect --format '{{.Id}}'.

These are the certified builds themselves, not a rebuild and not the public swebench/* images β€” the execution fingerprints of public swebench/* images do not match this set.

Stateful-tool domain

postgres:15 is a public Docker Hub image (docker pull postgres:15). The Toolathlon-Gym runtime and the MCP server set are the reproducible default for the tool domain (native live-API Toolathlon is optional and may incur cost).

Metrics

Two standard measures and two memory-specific measures. Infrastructure failures are reported separately and never counted as wrong answers.

Metric Definition
SR Task Success Rate β€” end-to-end workflow success. Reported as completed / registered.
VNCR Verified Native Completion Rate β€” macro-mean of native_passed over nodes: whether each node completed its own declared obligation under domain-native checks.
VHS Verified Handoff Success β€” micro-ratio of verified eligible handoffs, i.e. downstream success conditioned on successful mandatory predecessors. Proxy for required-memory availability.
ICS Isolation Challenge Score β€” success on polluted instances conditioned on matched control success.

ICS is the memory-boundary metric. It is a conditional rate: a polluted instance only counts toward the denominator when its matched control was solved. If the denominator is below 10, ICS is reported blank (β€”), never as 0.

Cost is reported as Tokens / Task (input + output, median and P90), covering agent + coordination + memory + recovery traffic.

What is intentionally not included

This repository is the task side of the benchmark β€” everything a system needs to see in order to run. The following are held out for scoring and are not here:

  • evaluation oracles (gold answers, gold patches, gold proofs, native evaluator inputs),
  • gold execution trajectories,
  • pollution plans (the mutation operator, target node, and expected isolation action for each polluted instance).

Container images are also outside this repository β€” see Execution environments above.

Systems can be executed and evaluated against the released workflows; final scores come from the held-out evaluation.

Loading

Plain JSON is enough:

import json

def load(path):
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f]

families  = load("data/workflow_families.jsonl")
contexts  = load("data/node_contexts.jsonl")
topology  = load("data/topology_descriptors.jsonl")
math      = load("data/math_cases.jsonl")
corpora   = load("data/retrieval_corpora.jsonl")

With HuggingFace datasets, point data_files at the repository's local files (or the paths served by your copy of this repository):

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "families": "data/workflow_families.jsonl",
        "node_contexts": "data/node_contexts.jsonl",
        "topology_descriptors": "data/topology_descriptors.jsonl",
        "math_cases": "data/math_cases.jsonl",
        "retrieval_corpora": "data/retrieval_corpora.jsonl",
        "control_instances": "data/control_instances.jsonl",
        "polluted_instances": "data/polluted_instances.jsonl",
        "benchmark_instances": "data/benchmark_instances.jsonl",
    },
)

Versioning

This is v1 of the public release. The 800 problems are complete and there is no held-out test/dev subdivision.

License

The annotations, workflow structures, and memory-context scopes in this repository are released under CC BY 4.0.

Problems are constructed from upstream sources; those upstream assets remain under their own terms:

Domain Upstream
stateful_tool_use Toolathlon
repository_code_change SWE-bench Verified
offline_information_retrieval BrowseComp+
mathematical_and_structured_reasoning Lean4 proof-dependent workflows

Please cite both CoMemBench and the upstream sources if you use the corresponding domains.

Citation

@misc{comembench2026,
  title        = {{CoMemBench}: Benchmarking Collaborative Memory
                  Boundaries across Multi-Agent Workflow Topologies},
  author       = {Anonymous},
  year         = {2026},
  note         = {Manuscript under review}
}
Downloads last month
32