Arc-Intelligence/arc-crm-benchmark — Dataset Adaptation

Source assessment

MCP / tool-use evidence: Present in inspected metadata
Professional domain: Sales Marketing
License: mit · within automated reuse policy
Assessment observed: 2026-09-04T21:31:25+00:00
Source revision: 8efb61f5920812d03d6267ed25f4439acb82cfbf

Inspect classification evidence

{
  "markers": [
    "tool-calling",
    "tool-use"
  ],
  "source": "catalog-metadata"
}

Metadata signals guide routing; they do not prove runtime fidelity or task quality.

The Blobfish adaptation

Arc CRM 6 v0.1.2 publishes 6 independent synthetic workflows. All 1,200 source conversations were inspected, but 0 source rows are reproduced. This is partial coverage, not upstream API parity.

Local Docker reference oracle: 6/6 strict passes
Registry Docker reference oracle: 6/6 strict passes
Model leaderboard: No model evaluation yet; oracle passes are not model scores.
Hugging Face commit: b53b22d65fa4a82d22fc95b91326b381b1482c44
Frozen manifest SHA256: 1daae8b6c5f34f33e014600cea82339fd5b4507c090af8bf08da37727b86e6a3
Harbor immutable identifier: sha256:fc9ca3b80d6a1a3c3e5bd1b975c54f1f3ada21b49668cb5f25e2912d67b69dfa

Tasks, assets & environment →Reference trajectories →

Pinned Hugging Face dataset ↗Harbor release ↗Executable GitHub world ↗

Model evaluation and further source coverage remain pending. The historical metadata assessment is retained separately from this publication-backed status.

Related working example

salesbench-100 provides a domain reference. It does not count as this dataset’s adaptation.

View tasks, assets, environment & trajectories →