Hugging Face's logo

Datasets:
Novylab
/
mssense-eval-benchmark

Tasks:
Text Generation
Modalities:
Text
Formats:
json
Languages:
English
French
Size:
1K - 10K
Tags:
rpa
robotic-process-automation
action-trace
closed-vocabulary
structured-generation
workflow
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
mssense-eval-benchmark / docs /datasheet.md
Novylab's picture
Novylab
Upload 13 files
b6a6d0d verified
preview code
|
Raw
History Blame Contribute Delete
10.9 kB

Datasheet for mssense Evaluation Benchmark

Following the Datasheet for Datasets template (Gebru et al., 2018, arXiv:1803.09010). Companion to the lighter dataset_card.md and the evaluation_protocol.md. Platform-specific terms (mssense, LAT, LeBrain, IRIS, Intentia) are defined in the Terminology table of the README.

Motivation

For what purpose was the dataset created? To provide an evaluation suite for mssense — the conversational intent-understanding and workflow-validation component developed within the Intentia research programme (Novelis R&D, 2026). Tasks: clarification policy, LAT audit, semantic judgment, workflow creation, business-rule extraction, visual grounding / governance and modification-intent handling. The benchmark covers the heterogeneity of inputs (text instructions, audio transcripts, screen captures, demo recordings) and outputs (UNDERSTOOD, VALID, INVALID, QUESTION, THINKING, DONE).

The systems it evaluates are two method variants of mssense: HCA-WA (Hybrid Checkpointed Agent for Workflow Authoring) and SGAR-WA (Skill-Grounded Agentic Reasoning for Workflow Authoring, the current default).

Who created the dataset and on behalf of which entity? Authored within the Intentia research programme (Novelis R&D). The mssense_public_creation_wonderbread_v1_full sub-corpus is derived from the public WONDERBREAD dataset (see Provenance in LICENSE-DATA).

Who funded the creation of the dataset? Internal R&D budget.

Composition

What do the instances comprising the dataset represent? Each instance represents one mssense-flavoured task: validate a LAT, audit a candidate LAT for ambiguity, generate a workflow from a description, extract a business rule from a free-form prompt, or judge a modification intent. The unit of an instance is one JSONL record carrying:

  • benchmark_version
  • input_payload (text, audio_url, capture artifact reference, demo reference)
  • lat_candidate (the LAT under analysis, when relevant)
  • business_rules (BR list, when relevant)
  • channel_family (web | desktop | other)
  • input_modality (text | audio | capture | demo)
  • expected_decision (VALID | INVALID | UNDERSTOOD | QUESTION | THINKING | DONE)
  • expected_audit_question (when expected_decision == INVALID)
  • expected_issue_types (list — AMBIGUOUS_SELECTOR, MISSING_PRECONDITION, UNRESOLVED_VARIABLE, INCONSISTENT_FLOW, MISSING_VALUE)
  • expected_choices
  • oracle (machine-readable ground truth)
  • provenance (block, origin_sample_id, source)
  • difficulty, content_hash

How many instances are there in total?

No train/dev/test split.

Does the dataset contain all possible instances or is it a sample? A curated sample union of three sub-corpora:

  • mssense_public_audit_internal_v1* — internal-origin audit cases
  • mssense_public_validation_internal_v1* — internal-origin validation cases
  • mssense_public_creation_wonderbread_v1_full* — WONDERBREAD-derived creation/demo cases

What data does each instance consist of? One JSONL record in data/mssense_eval_benchmark_v1_1.jsonl, schema validated by schema/evaluation_sample.v1_1.schema.json. Cross-references to upstream WONDERBREAD ids are preserved in provenance.origin_sample_id when applicable. No PII; all names are synthetic.

Are there recommended data splits? No. Pure evaluation suite. Reports are stratified by task_family, channel_family, input_modality, expected_decision and difficulty.

Are there any errors, sources of noise, or redundancies in the dataset?

  • Multiple expected_decision codes coexist; the most frequent are INVALID, VALID, UNDERSTOOD, QUESTION. Methods that emit only a subset of these labels must document how they normalize.
  • WONDERBREAD-derived samples carry a more permissive ambiguity profile than internal-origin samples; stratified reporting captures this.

Is the dataset self-contained, or does it link to or otherwise rely on external resources? Self-contained for the text / metadata path. Audio / capture / demo modalities reference relative artifact paths under data/; the original WONDERBREAD videos and key frames are not republished here (license-restricted) and must be obtained from the upstream dataset if needed for grounding studies.

Collection process

How was the data acquired?

  • Internal audit / validation sub-corpora: authored in-house by domain experts using a structured authoring rubric.
  • WONDERBREAD-derived creation sub-corpus: extracted programmatically by an internal pipeline that reads upstream metadata.json + intent + action_trace + SOP and emits records conforming to the canonical schema (schema/evaluation_sample.v1_1.schema.json).

What mechanisms or procedures were used to collect the data?

  • Internal authoring tools producing JSONL.
  • A Python extractor for the WONDERBREAD-derived sub-corpus (internal, not distributed).
  • Manual review of hard cases against a publication-readiness checklist.

Over what timeframe was the data collected? Authoring + extraction: April–June 2026.

Were any ethical review processes conducted? Not formally. Internal review focused on (a) absence of PII in audit/validation cases, (b) compliance with WONDERBREAD's upstream license for the creation sub-corpus, (c) no real customer data exposed.

Preprocessing / cleaning / labeling

Was any preprocessing/cleaning/labeling done?

  • Label canonicalisation: upstream WONDERBREAD intent is mapped to the mssense expected-decision taxonomy.
  • LAT extraction: action_trace is normalised into the LAT shape ({id, channel, action, params, …}).
  • Content hashing per record (content_hash).
  • Hard-case adjudication: a subset of high-difficulty WONDERBREAD samples was manually reviewed for label correctness.

Was the "raw" data saved in addition to the preprocessed/cleaned/labeled data? The extractor inputs and intermediate manifests are retained internally for reproducibility; they are not vendored here. The original WONDERBREAD media is not republished either (see LICENSE-DATA).

Uses

Has the dataset been used for any tasks already?

  • Internal mssense regression evaluation across successive method versions.
  • End-to-end approach sweeps over the mssense HTTP surface.

Is there a repository that links to any or all papers or systems that use the dataset? This dataset is the evaluation suite associated with the manuscript Closed-Vocabulary Action Trace Generation for Conversational RPA Authoring (Yahaya Alassan, Ettifouri, Dahhane; Novelis), submitted to the Journal of Object Technology.

What (other) tasks could the dataset be used for? Any evaluation of clarification-policy agents, structured-question generation, LAT auditing methods, or workflow-creation systems on a multi-modal corpus.

Are there tasks for which the dataset should not be used?

  • Training a model from scratch (evaluation-only by design).
  • Republishing WONDERBREAD images / videos via this benchmark (the original upstream license governs those artifacts).

Why a New Benchmark? (V5.15.h.3)

A reviewer of this dataset will reasonably ask whether an existing public corpus could substitute for it. The detailed answer is in related_benchmarks_comparison.md (twelve candidate public datasets analysed property-by-property) and why_new_benchmark.md (one-page gap analysis). In short: the problem requires four properties — (i) closed-vocabulary RPA actions, (ii) conversational input, (iii) multi-step durable trace, (iv) multi-modal coverage — and no public corpus satisfies all four simultaneously. WONDERBREAD is reused as the upstream source of the workflow_creation sub-corpus but cannot substitute for the whole.

Statistical power and inter-annotator agreement (V5.15.h.3)

  • A priori statistical-power analysis per research question is in statistical_power_analysis.md. The full corpus supports detection of 5–10 % absolute effect-size differences at α = 0.05, power = 0.80, on the four RQ-relevant comparisons.
  • Inter-annotator agreement is not applicable to 1593 of the 1865 samples (the controlled-balanced and augmented blocks) by construction; the remaining 272 internally-authored seeds followed a documented rubric with hard-case team review. See inter_annotator_agreement.md. A formal κ measurement on a stratified subset is planned for the v1.2 release.

Distribution

Will the dataset be distributed to third parties outside of the entity on behalf of which it was created? Yes for the internal-origin sub-corpora under CC-BY-4.0. Distribution of the WONDERBREAD-derived sub-corpus is conditioned on attribution to WONDERBREAD upstream.

How will the dataset be distributed? JSONL + schemas + harness + reports + manifest, with the LICENSE-DATA provenance clause attached. Hosting channel TBD.

Will the dataset be distributed under a copyright or other intellectual property (IP) license?

  • Code, scripts, schema: Apache-2.0 (LICENSE).
  • Internal-origin data: CC-BY-4.0 (LICENSE-DATA).
  • WONDERBREAD-derived sub-corpus: CC-BY-4.0 + cumulative attribution to WONDERBREAD upstream.

Have any third parties imposed IP-based or other restrictions on the data associated with the instances? The WONDERBREAD upstream license imposes attribution; this benchmark complies by (a) explicit citation in LICENSE-DATA, (b) not republishing upstream media.

Maintenance

Who will be supporting/hosting/maintaining the dataset? The Intentia research programme (Novelis R&D). The dataset is hosted at its DOI landing page (https://doi.org/10.5281/zenodo.21105006).

Will the dataset be updated? Yes. v1.0-eval and v1.1-eval are the first published milestones; changes are tracked in ../reports/CHANGELOG_v1.0_to_v1.1.md.

Will older versions of the dataset continue to be supported/hosted/maintained? Yes — each versioned release is preserved under its own DOI on the hosting platform.

If others want to extend/augment/build on/contribute to the dataset, is there a mechanism for them to do so? Contributions are welcome via the dataset's DOI landing page. Extensions must respect the canonical schema (schema/evaluation_sample.v1_1.schema.json) and document any new label codes.