RobustOps
Collection
4B enterprise agents: task completion, fault recovery, models and trajectories. Code: https://github.com/haohuiliang/RobustOps • 4 items • Updated • 1
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/harbor/harbor.py", line 171, in _split_generators
raise DataFilesNotFoundError("No task.toml or instruction.md files found")
datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
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/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
RobustOps-Trajectories
English · 简体中文 · Models & data collection
GitHub: code & reproduction · Quick start · Training · Evaluation
RobustOps combines supervised fine-tuning and reinforcement learning to improve task completion and fault recovery in 4B enterprise tool agents. The GitHub project provides source code, enterprise runtime adapters, fault injection, training data, recipes and evaluation workflows.
Optional trajectory evidence for reviewing RobustOps evaluation runs. Training data is included in the GitHub repository; model training and evaluation do not require this archive.
Collections
The archive contains 5,507 compact trajectory records, including 760 final Test attempts, with model/split/temperature folders preserved. Oracle exposes task tools; plus5 adds five distractor tools.
| Collection | Temperature | Sampling | Records |
|---|---|---|---|
| Base clean Val/Test | 0 | Final scored branch per case | 127 / 130 |
| Base fault Val/Test | 0 | Final scored branch per case | 127 / 130 |
| Guided teacher Train | 0 / 0.6 | Initial collection and failure retry | 2,701 |
| Guided teacher Val | 0 | Accepted supervised-loss diagnostics | 15 |
| Unguided teacher Test | 0 | Final clean / actual-hit fault | 130 / 124 |
| SFT Val | 0 / 0.6 | Single sample / four samples per clean/fault condition | 254 / 1,016 |
| SFT Test | 0 | Final clean / actual-hit fault | 130 / 121 |
| RL120 Val | 0 | Final clean / actual-hit fault | 127 / 120 |
| RL120 Test | 0 | Final clean / actual-hit fault | 130 / 125 |
Download and inspect
Download robustops-trajectories.tar.gz. Extract it from the RobustOps source root; the resulting trajectories/ directory is ignored by Git. The separate manifest.json lists collections and final Test selections. Full folder descriptions are inside the archive.
# Download the optional trajectory archive from the RobustOps source root.
hf download haohuiliang/RobustOps-Trajectories robustops-trajectories.tar.gz \
--repo-type dataset --local-dir dist/hf
# Extract the recorded evidence into the local trajectories/ directory.
tar -xzf dist/hf/robustops-trajectories.tar.gz
Retained evidence
Records preserve executed assistant/tool histories, fault events, recorded verifier values, provenance and collection indexes. Guided teacher sources retain the final logged request with conversation and tool definitions for supervision processing. The processed 692 Train and 15 Val diagnostic records are included in the source repository.
Final Test and RL Val fault selection keeps the original valid hit, otherwise the first valid hit among at most four supplements, including business failures. Invalid collections, superseded attempts and zero-hit faults are removed. SFT/RL/teacher have 9/5/6 no-hit Test items; the evaluation denominator remains 130. Actual-hit task success differs from recovery, which also requires matched clean success.
SFT Pass@4 uses four temperature-0.6 attempts to inspect available successful paths before RL. Test evidence does not enter training.
Reproduction and limits
The compact archive supports review and recorded-result recomputation. It omits database snapshots, duplicate branches and other per-turn request/response files. Recorded verifier values support comparison checks; independent historical SQL execution would require terminal databases, which are not included.
Tasks and synthetic state derive from EnterpriseOps-Gym, dataset revision c8e538eae8a6205294f0a86675fefdc1fac408f6. Evaluation details
License and attribution
Apache-2.0 with upstream terms and attribution. Preserve LICENSE, THIRD_PARTY_NOTICES.md and teacher provenance when redistributing. EnterpriseOps-Gym · Arctic-MidTool-RL-4B · veRL.
- Downloads last month
- 407