Novylab/mssense-eval-benchmark

mssense Evaluation Benchmark — Closed-Vocabulary Action Trace Generation

Canonical version / DOI: archived on Zenodo at https://doi.org/10.5281/zenodo.21105006 (CC-BY-4.0). This Hugging Face repository is a distribution mirror — please cite the Zenodo DOI.

An evaluation-only benchmark for closed-vocabulary action trace generation in conversational Robotic Process Automation (RPA) authoring. Each sample pairs a conversational request with the oracle labels needed to judge whether a generated action trace is executable against a closed, typed, channel-specific action catalogue — not merely schema-valid.

• Version: 1.1-eval
• Samples: 1865 (1772 seeds + 93 deterministic paraphrastic variants)
• Task families (9): clarification policy, LAT audit, semantic judgment, workflow creation, business-rule extraction, visual grounding / governance, modification intent, audit, interaction regression
• Split: none — the full file is the evaluation suite
• License: Creative Commons Attribution 4.0 International (CC-BY-4.0)

Terminology

A few names in this benchmark are specific to the platform it originates from. They are kept verbatim because they are used throughout the samples and schema (for example, in sample_id prefixes and iris_* field names) and changing them would break reproducibility and the dataset's published identity. The acronyms are platform-specific; the problems they instantiate are general and platform-independent.

Term	Meaning (community-standard concept)
action trace	an ordered sequence of typed, executable actions — the durable artefact the system must produce
LAT (LeBrain Action Trace)	the platform-specific instance of an action trace used in this benchmark; a list of typed steps. The field lat_candidate holds the candidate trace under analysis
mssense	the conversational intent-understanding and workflow-validation component evaluated by this benchmark (the system under test); also the benchmark's name
LeBrain	the automation / intelligence platform (Novelis) that connects applications and automates business processes; it defines the closed action catalogue and executes the traces
IRIS	LeBrain's Computer Use Agent; the iris_* fields (e.g., iris_control_type) describe executable UI steps targeted at IRIS
Intentia	the 2026 research programme of the Novelis R&D laboratory, within which mssense is developed
channel	an action category / connector — web, desktop, spreadsheet, email, database, API, file, control-flow
oracle	the per-sample ground-truth block (expected_decision, expected_issue_types, required_checks) used for scoring

Contents

data/mssense_eval_benchmark_v1_1.jsonl     the benchmark (one JSON object per line)
schema/evaluation_sample.v1_1.schema.json  JSON Schema for a sample
docs/datasheet.md                          Datasheet for Datasets (Gebru et al., 2018)
docs/dataset_card.md                       dataset card
docs/evaluation_protocol.md                metrics, splits, scoring conventions
docs/related_benchmarks_comparison.md      property-by-property comparison of 16 public benchmarks
docs/why_new_benchmark.md                  one-page gap analysis
docs/statistical_power_analysis.md         a priori power analysis per research question
docs/inter_annotator_agreement.md          IAA disclosure and v1.2 roadmap
reports/CHANGELOG_v1.0_to_v1.1.md          changes from v1.0 to v1.1
LICENSE-DATA                               CC-BY-4.0
CITATION.cff                               citation metadata
SHA256SUMS.txt                             integrity checksums

Verify integrity with sha256sum -c SHA256SUMS.txt (or certutil -hashfile <file> SHA256 on Windows).

Sample format

One JSON object per line. Key fields include sample_id, task_family, channel_family, input_modality, difficulty, user_intent, input_payload, lat_candidate, expected_decision, expected_issue_types, business_rules, and an oracle object with required_checks. See schema/evaluation_sample.v1_1.schema.json and docs/datasheet.md for the full specification.

The issue-type vocabulary is the platform's canonical set: MISSING_VALUE, UNRESOLVED_VARIABLE, AMBIGUOUS_SELECTOR, MISSING_PRECONDITION, INCONSISTENT_FLOW.

import json
samples = [json.loads(l) for l in open("data/mssense_eval_benchmark_v1_1.jsonl", encoding="utf-8")]
print(len(samples), "samples")

Provenance and license

The released benchmark comprises internally-authored audit, validation, and generation cases, licensed under CC-BY-4.0. A WONDERBREAD-derived sub-corpus is prepared in the release tree under a forward-looking attribution clause and is not included in this v1.1 evaluation file pending adjudication; the clause takes effect on its first integrated release. Full provenance is documented in docs/datasheet.md.

Privacy and sanitization. This public release is privacy-sanitized: internal authoring paths in the input_payload.source_file field were reduced to file basenames, and incidental personal data that appeared as example form-fill values in a few interaction scenarios were replaced with synthetic values. These are metadata and scenario-input fields only; no oracle label (expected_decision, expected_issue_types, oracle) was modified, so the evaluation is unaffected.

Associated publication

This benchmark supports the manuscript Closed-Vocabulary Action Trace Generation for Conversational RPA Authoring (Yahaya Alassan, Ettifouri, Dahhane; Novelis), submitted to the Journal of Object Technology.

Citation

Dataset DOI: https://doi.org/10.5281/zenodo.21105006 (CC-BY-4.0). See CITATION.cff for machine-readable citation metadata.