Datasets:

MAS-Sys
/

MAS-Sys-Traces

License:

Dataset Viewer

Split (2)

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
pytest-dev__pytest-7168: struct<patch_is_None: bool, patch_exists: bool, patch_successfully_applied: bool, resolved: bool, te (... 325 chars omitted)
  child 0, patch_is_None: bool
  child 1, patch_exists: bool
  child 2, patch_successfully_applied: bool
  child 3, resolved: bool
  child 4, tests_status: struct<FAIL_TO_PASS: struct<success: list<item: string>, failure: list<item: null>>, PASS_TO_PASS: s (... 212 chars omitted)
      child 0, FAIL_TO_PASS: struct<success: list<item: string>, failure: list<item: null>>
          child 0, success: list<item: string>
              child 0, item: string
          child 1, failure: list<item: null>
              child 0, item: null
      child 1, PASS_TO_PASS: struct<success: list<item: null>, failure: list<item: null>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: null>
              child 0, item: null
      child 2, FAIL_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: null>
              child 0, item: null
      child 3, PASS_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: null>
              child 0, item: null
pylint-dev__pylint-7114: struct<patch_is_None: bool, patch_exists: bool, patch_successfully_applied: bool, resolved: bool, te (... 329 chars omitted)
  child 0, patch_is_None: bool
  child 1, patch_exists: bool
  child 2, patch_successfully_applied: bool
  child 3, resolved: bool
  child 4, tests_status: struct<FAIL_TO_PASS: struct<success: list<item: null>, failure: list<item: string>>, PASS_TO_PASS: s (... 216 chars omitted)
      child 0, FAIL_TO_PASS: struct<success: list<item: null>, failure: list<item: string>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: string>
              child 0, item: string
      child 1, PASS_TO_PASS: struct<success: list<item: string>, failure: list<item: string>>
          child 0, success: list<item: string>
              child 0, item: string
          child 1, failure: list<item: string>
              child 0, item: string
      child 2, FAIL_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: null>
              child 0, item: null
      child 3, PASS_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
          child 0, success: list<item: null>
              child 0, item: null
          child 1, failure: list<item: null>
              child 0, item: null
to
{'pylint-dev__pylint-7114': {'patch_is_None': Value('bool'), 'patch_exists': Value('bool'), 'patch_successfully_applied': Value('bool'), 'resolved': Value('bool'), 'tests_status': {'FAIL_TO_PASS': {'success': List(Value('null')), 'failure': List(Value('string'))}, 'PASS_TO_PASS': {'success': List(Value('string')), 'failure': List(Value('string'))}, 'FAIL_TO_FAIL': {'success': List(Value('null')), 'failure': List(Value('null'))}, 'PASS_TO_FAIL': {'success': List(Value('null')), 'failure': List(Value('null'))}}}}
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
              pytest-dev__pytest-7168: struct<patch_is_None: bool, patch_exists: bool, patch_successfully_applied: bool, resolved: bool, te (... 325 chars omitted)
                child 0, patch_is_None: bool
                child 1, patch_exists: bool
                child 2, patch_successfully_applied: bool
                child 3, resolved: bool
                child 4, tests_status: struct<FAIL_TO_PASS: struct<success: list<item: string>, failure: list<item: null>>, PASS_TO_PASS: s (... 212 chars omitted)
                    child 0, FAIL_TO_PASS: struct<success: list<item: string>, failure: list<item: null>>
                        child 0, success: list<item: string>
                            child 0, item: string
                        child 1, failure: list<item: null>
                            child 0, item: null
                    child 1, PASS_TO_PASS: struct<success: list<item: null>, failure: list<item: null>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: null>
                            child 0, item: null
                    child 2, FAIL_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: null>
                            child 0, item: null
                    child 3, PASS_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: null>
                            child 0, item: null
              pylint-dev__pylint-7114: struct<patch_is_None: bool, patch_exists: bool, patch_successfully_applied: bool, resolved: bool, te (... 329 chars omitted)
                child 0, patch_is_None: bool
                child 1, patch_exists: bool
                child 2, patch_successfully_applied: bool
                child 3, resolved: bool
                child 4, tests_status: struct<FAIL_TO_PASS: struct<success: list<item: null>, failure: list<item: string>>, PASS_TO_PASS: s (... 216 chars omitted)
                    child 0, FAIL_TO_PASS: struct<success: list<item: null>, failure: list<item: string>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: string>
                            child 0, item: string
                    child 1, PASS_TO_PASS: struct<success: list<item: string>, failure: list<item: string>>
                        child 0, success: list<item: string>
                            child 0, item: string
                        child 1, failure: list<item: string>
                            child 0, item: string
                    child 2, FAIL_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: null>
                            child 0, item: null
                    child 3, PASS_TO_FAIL: struct<success: list<item: null>, failure: list<item: null>>
                        child 0, success: list<item: null>
                            child 0, item: null
                        child 1, failure: list<item: null>
                            child 0, item: null
              to
              {'pylint-dev__pylint-7114': {'patch_is_None': Value('bool'), 'patch_exists': Value('bool'), 'patch_successfully_applied': Value('bool'), 'resolved': Value('bool'), 'tests_status': {'FAIL_TO_PASS': {'success': List(Value('null')), 'failure': List(Value('string'))}, 'PASS_TO_PASS': {'success': List(Value('string')), 'failure': List(Value('string'))}, 'FAIL_TO_FAIL': {'success': List(Value('null')), 'failure': List(Value('null'))}, 'PASS_TO_FAIL': {'success': List(Value('null')), 'failure': List(Value('null'))}}}}
              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.

MAS-Sys Main-Experiment Traces

This dataset contains the system traces and post-hoc task-quality results in the MAS-Sys main experiment.

Experiment matrix

Hardware	Target JPS values	Backends	Cells
2x H200	0.005, 0.01, 0.02, 0.05, 0.1	5	25
4x H200	0.01, 0.02, 0.05, 0.1	5	20

The five backend configurations are continuum, vllm-omni, vllm, vllm-omni, sglang, orla+vllm, vllm-omni, and orla+sglang. Each cell uses a deterministic mixture of ResearchBench, SWE-bench, DeepFundBench, PinchBench, and PaperBananaBench. All cells use seed 42, a 30-minute arrival window, and a 1,200-second per-job deadline. Backends at the same hardware/load point share the same arrival plan.

Dataset contents

.
├── README.md
├── LICENSE
├── runs.csv
├── quality.csv
└── raw/
    └── <run_id>/
        ├── config_snapshot.json
        ├── runtime_assets.json
        ├── overall_result.json
        ├── job_results.json
        ├── resource_metrics.json
        ├── backend_health.json
        ├── jobs/<job_id>/
        └── posthoc_artifacts/

Both CSV files use two header rows and contain one row per experiment cell. Their run_id field maps directly to raw/<run_id>/.

runs.csv contains system-level measurements, including terminal-state counts, runtime, token and tool statistics, inference latency, concurrency, resource usage, backend health, and ORLA timing. It does not use task quality to define execution success.
quality.csv contains submitted/evaluated job counts and effective and evaluated-only task-quality results. The task metrics are SWE-bench resolved, ResearchBench embedding_sim, DeepFundBench direction score (profit/flat/loss mapped to 1/0.5/0), PinchBench combined_score, and PaperBananaBench overall_score / 10.

Raw records

Scope	File	Contents
Run	`config_snapshot.json`	Workload, model/backend configuration, and arrival plan
Run	`runtime_assets.json`	Runtime assets such as SWE-bench container images
Run	`overall_result.json`	Run-level aggregate results
Run	`job_results.json`	Compact index of all jobs in the run
Run	`resource_metrics.json`	Timestamped CPU, GPU, and KV-cache samples
Run	`backend_health.json`	Timestamped serving-endpoint health samples
Job	`result.json`	Terminal state and benchmark-specific output
Job	`posthoc_evaluation.json`	Benchmark-specific quality evaluation
Job	`trace.json`	Full MAS execution record, including planning, agent, tool, and LLM events
Job	`metadata.json`	Structural and timing summary of the execution
Job	`llm_calls.json`	Flattened per-call serving and latency measurements

Deadline-terminated jobs may use deadline_debug_trace.json, deadline_debug_metadata.json, and deadline_debug_llm_calls.json to preserve the last available execution state. DeepFundBench jobs can additionally contain date-level result, trace, metadata, and LLM-call files. SWE-bench post-hoc artifacts include predictions, patches, test logs, and evaluation reports.

Per-job Docker work directories, byte-identical deadline-debug copies, and reproducible diagnostic plots are not included. Base64-encoded images are retained when they were part of a recorded model input or output.

Interpretation notes

succeeded_jobs, failed_jobs, and killed_jobs are workflow terminal states, not task-correctness labels.
Failed and partial traces are retained because failure behavior is part of the system-level evidence.
effective_mean_score assigns zero to unevaluated submissions, whereas evaluated_only_mean_score conditions on completed evaluations. Both should be read together with the submitted/evaluated job counts.
A degraded health label in an SGLang-family run can reflect a busy health endpoint even when inference requests continue to complete.
Some ORLA traces preserve observed structured-history flattening and the resulting empty_patch outcomes; these records were not removed post hoc.
Run 20260505_182006 contains a historical deadline race in its raw summary. runs.csv reports mutually exclusive terminal states with success taking precedence, while the original raw evidence is retained.

Privacy and reproducibility

Experiment-operator home-directory paths were replaced with <MAS_SYS_HOME>. Public names and paths originating in benchmark inputs are preserved as task content.

The release records workload configurations, arrival plans, backend/model labels, execution traces, resource and health samples, and post-hoc evaluation artifacts. Historical runs do not consistently capture exact Git states, container digests, model revisions, dependency lockfiles, or every evaluator version; the data should therefore not be interpreted as a byte-for-byte execution-environment snapshot.

License

Unless otherwise indicated, the MAS-Sys experimental measurements, trace organization, annotations, and documentation in this dataset are released under the Apache License 2.0.

Downloads last month: 118

Total file size:

8.39 GB