yasserrmd/enterprise-reflex-dataset

Enterprise Reflex Dataset

Repository: yasserrmd/enterprise-reflex-dataset
Creator: Mohamed Yasser
Language: English
Task: Enterprise action selection, dynamic action ranking, abstention, and System-2 routing
Status: V0 Research Dataset

Dataset Summary

Enterprise Reflex Dataset is a structured dataset for training lightweight enterprise decision models that operate as a fast System-1 layer in front of larger reasoning models.

The dataset is designed around a different problem from conventional intent classification.

Instead of predicting from a fixed list of labels, a model receives:

• a user or system request
• optional enterprise state
• optional contextual information
• a runtime-defined set of candidate actions

The model must determine which action best satisfies the request or select NO_ACTION when none of the available actions are appropriate.

The dataset supports the development of models capable of:

• dynamic action ranking
• enterprise request-to-action matching
• same-domain action discrimination
• abstention
• tool selection
• workflow routing
• System-1 / System-2 architectures

Motivation

Enterprise AI systems increasingly expose large numbers of APIs, tools, workflows, services, and agent actions.

Sending every request directly to a large reasoning model can introduce:

• unnecessary latency
• higher inference cost
• larger tool-selection spaces
• inconsistent routing
• unnecessary reasoning for simple decisions

Enterprise Reflex explores an alternative architecture:

Request + State + Candidate Actions
               |
               v
      Enterprise Reflex
               |
        Action Ranking
               |
      +--------+--------+
      |                 |
  Confident          Uncertain
      |                 |
    Action          System-2 LLM

The dataset was created to support this type of selective decision layer.

Dataset Configurations

The repository contains two primary configurations.

canonical

The reusable high-level representation of enterprise decision tasks.

Typical structure:

{
  "id": "example-id",
  "language": "en",
  "domain": "procurement",
  "subdomain": "requisition",
  "request": "Create a new requisition for approved laptops.",
  "state": {
    "approval_status": "approved"
  },
  "context": {},
  "candidate_actions": [
    {
      "name": "procurement.create_requisition",
      "description": "Create a new procurement requisition.",
      "family": "CREATE",
      "domain": "procurement"
    }
  ],
  "gold_actions": [
    "procurement.create_requisition"
  ],
  "no_action": false,
  "metadata": {},
  "provenance": {
    "source_dataset": "...",
    "source_repo": "...",
    "source_license": "...",
    "source_url": "...",
    "source_row_id": "..."
  }
}

pairs

Pairwise representation used to train Enterprise Reflex V0.

Each request/state representation is paired with one candidate action and labelled according to whether that candidate is appropriate.

Conceptually:

(request + state + context, candidate action)
                    |
                    v
                 0 / 1

This allows candidate actions to be supplied dynamically instead of requiring a fixed classification head for every possible enterprise action.

Data Construction

The dataset combines normalized enterprise-oriented tasks and tool-use examples from multiple upstream datasets.

Sources used during V0 construction include:

• ODE Enterprise Use Cases
• ServiceNow EnterpriseOps-Gym
• Enterprise Worlds
• ServiceNow task datasets
• APIGen function-calling data
• IT Support V2

Some datasets investigated during development were intentionally excluded from the primary training set because of access, redistribution, or licensing constraints.

Examples include:

• WorkArena instance data
• CRMArena-Pro by default
• gated EnterpriseBench data as training material

Action Ontology

Actions are represented using human-readable names and descriptions rather than fixed numeric class IDs.

Examples:

procurement.create_requisition
procurement.create_purchase_order
procurement.request_approval

itsm.create_incident
itsm.create_problem
itsm.create_change

hr.update_employee
hr.request_leave

iam.reset_password

Broad action families include:

CREATE
UPDATE
DELETE
SEARCH
NOTIFY
ROUTE
EXECUTE
OTHER
ABSTAIN

NO_ACTION represents explicit abstention.

NO_ACTION Generation

NO_ACTION examples are not intended to mean simply "unsafe request."

A major V0 construction strategy was to take an otherwise valid enterprise task and remove its valid action from the available candidate set.

The correct answer therefore becomes:

NO_ACTION

because none of the currently available actions correctly satisfy the request.

This allows the model to learn that it is not required to select a tool merely because tools are available.

Hard Negatives

The dataset contains negative candidate actions to teach action discrimination.

V0 contains both:

• cross-domain negatives
• same-domain negatives

Evaluation of V0 showed that future versions should contain a substantially higher proportion of difficult same-domain sibling actions.

Examples:

procurement.create_requisition
procurement.update_requisition
procurement.create_purchase_order
procurement.request_approval

and:

itsm.create_incident
itsm.create_problem
itsm.update_incident
itsm.close_incident

Splits

The dataset provides:

train
validation
test

Splitting is performed before pair expansion where possible to reduce leakage between related examples.

Current Research Findings

V0 evaluation indicates that the dataset is sufficient to train a strong request/action compatibility model.

However, harder evaluation exposed areas where future dataset versions should improve:

• same-domain sibling actions
• counterfactual state pairs
• enterprise policy constraints
• semantic cross-domain collisions
• calibrated abstention boundaries

Future versions should place substantially greater emphasis on these cases rather than simply increasing the total number of generic examples.

Recommended V1 Data Distribution

A future hard-training extension may emphasize approximately:

35%  same-domain sibling discrimination
25%  counterfactual state examples
15%  policy and workflow constraints
10%  semantic cross-domain collisions
8-10% NO_ACTION
5-7% ordinary/easier cases

These are design targets, not properties of the current V0 dataset.

Intended Uses

Appropriate research uses include:

• enterprise action routing
• agent tool selection
• API selection
• workflow routing
• lightweight System-1 decision models
• LLM tool-space reduction
• abstention research
• selective prediction
• enterprise agent architectures

Out-of-Scope Uses

The dataset should not be interpreted as sufficient evidence for autonomous execution of high-impact enterprise operations.

The dataset is not designed to independently authorize:

• financial transactions
• employee termination
• security administration
• destructive infrastructure operations
• legal decisions
• regulatory decisions
• safety-critical actions

Authorization should remain the responsibility of deterministic policy, workflow, access-control, and human-governance systems.

Limitations

English-first

V0 is primarily English.

Arabic and broader multilingual support are planned for later versions.

Synthetic Transformation

Some candidate sets, negative examples, normalized action descriptions, and NO_ACTION variants are generated or transformed from upstream tasks.

They may not reproduce the exact distribution of a live enterprise environment.

Domain Coverage

Although broader than a single enterprise domain, the dataset does not represent every possible enterprise workflow, application, policy, or state transition.

State Reasoning

Current evaluation suggests models trained on V0 can rely more strongly on request/action semantics than on subtle changes in structured enterprise state.

Future versions should contain substantially more counterfactual state pairs.

NO_ACTION

V0 models may over-predict abstention in some unfamiliar or weakly represented action areas.

Future versions should improve the boundary between:

valid but unfamiliar action

and:

no valid action available

Licensing and Provenance

This dataset was constructed from multiple upstream datasets with different licenses.

Provenance information is retained where possible at the example level.

Users are responsible for reviewing and complying with the licenses and terms associated with the original source datasets.

The presence of derived data in this repository should not be interpreted as replacing or overriding upstream licenses.

Known upstream sources used during construction include datasets released under licenses such as:

• Apache-2.0
• MIT
• CC-BY-4.0

Some investigated sources with restrictive or non-commercial terms were excluded from the default training pipeline.

Reproducibility

The pairs configuration preserves the training representation used for Enterprise Reflex V0.

The canonical configuration is intended to remain the reusable representation for future dataset and model iterations.

Version

Enterprise Reflex Dataset V0
September 2026

Citation

If this dataset is useful in your research or experiments, please reference the repository and creator:

@misc{yasser2026enterprisereflexdataset,
  author = {Mohamed Yasser},
  title = {Enterprise Reflex Dataset},
  year = {2026},
  publisher = {Hugging Face},
  note = {Enterprise action selection and System-1 decision dataset}
}

Disclaimer

This is an experimental research dataset.

Performance measured on this dataset or its associated benchmarks should not be interpreted as guaranteed performance in production enterprise environments.