Crmarenapro

[sampatsk]

Description

Adds the CRMArena-Pro dataset: ADP trajectories from Salesforce AI Research's
CRMArena-Pro benchmark, covering B2B and B2C Salesforce orgs. Rollouts are
produced by CRMArena's ReAct agent (--agent_strategy react, --org_type b2b|b2c)
and converted to ATIF. Assistant <thought>/<execute>/<respond> turns map to
agent steps; <execute> (SOQL/SOSL) normalizes to the built-in python tool and
respond stays a custom tool. Generated with GPT-5.1.

Type of Change

• New dataset

Dataset/Agent Information

• Name: crmarenapro
• Source: https://github.com/SalesforceAIResearch/CRMArena ; https://huggingface.co/datasets/Salesforce/CRMArenaPro
• Size: 12 representative sample trajectories (6 task categories × b2b/b2c)
• Domain: Salesforce CRM — service, sales, analytics, knowledge, confidentiality

Testing

• All tests pass (pytest -k crmarenapro; full suite with PYTHONPATH=$(pwd))
• Validation scripts run successfully
• Sample files generated correctly

Checklist

• Code follows style guidelines (ruff, mypy, pre-commit)
• Documentation is complete (README)
• All required files are present
• No sensitive data included