salesbench-100 / README.md
SamuelChien821's picture
Publish salesbench-100 v2 realism release
266027d verified
|
Raw
History Blame Contribute Delete
3.2 kB
metadata
license: cc-by-4.0
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - sales
  - salesforce
  - hubspot
  - gong
  - agents
  - mcp
  - deterministic-evaluation
pretty_name: SalesBench-100
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/tasks.jsonl

SalesBench-100

SalesBench-100 is a synthetic long-horizon sales-agent benchmark with 100 original workflows across Salesforce, HubSpot, Gong, and a seeded evidence room. Each task begins with a high-level employee request; the evidence room, CRM state, and embedded operating contract reveal the necessary investigation. Every task has 96 production-style source records, 281 task-specific deterministic criteria, 12 authorized CRM mutations, and its own accepted 163-call MCP trajectory.

Public release

Contents

  • data/tasks.jsonl: portable task records with prompts, context paths, deterministic rubric metadata, and public gold output.
  • tasks/: one readable JSON record per workflow.
  • task_files/: 9,600 unique seeded artifacts across Markdown, text, JSON, CSV, email, XML, and HTML.
  • world/: the four-surface offline MCP world and verifier.
  • contracts/: pinned contract metadata and exact published tool schemas.
  • trajectories/: accepted tool traces produced by the qualification run.
  • reports/: build, qualification, MCP-conformance, and container-probe evidence.

Objective release gates

Gate Required
Tasks 100
Workflow families 10
Documents per task 96
Authorized mutations per task 12
Reference MCP calls per task 163
Unique tool-name sequences 100/100
Task-specific deterministic criteria 281 per task
Verifier network/model/clock/random calls 0
Oracle and replay passes 100/100 each
False accepts across six negative controls 0
Reward for the pristine no-op control exactly 0.0

Measured results live in reports/qualification.json. Reference traces are implementation proofs, not model scores.

Contract fidelity

The world keeps Salesforce, HubSpot, Gong, and filesystem as separate MCP endpoints. Names and input JSON Schemas are pinned to official hosted documentation or real open-source implementations at immutable commits recorded in contracts/upstream-pins.json. The Gong surface is read-only. Output envelopes mirror the corresponding MCP/API JSON shapes, with documented offline simplifications.

Data and licensing

All people, entities, records, messages, amounts, and commercial facts are synthetic. CRMArena, SCUBA, AutomationBench, Harvey Labs, and Apex Accounting informed quality and packaging analysis only; their task text and data are not included. Synthetic task data is CC-BY-4.0; generator/runtime code is Apache-2.0. Third-party contract facts retain their upstream terms.