Skip to content

Repository files navigation

SalesTranscriptQA

A dialogue RAG benchmark built from Salesforce's synthetic CRMArena-Pro sales calls, with separate B2B and B2C question sets and a companion CLI in this repository.

Status: 200-question pilot complete. All 200 questions passed automated validation: 50 single-call and 50 two-call questions per domain, with the full 10,829-call corpus. No human review or calibration was performed. The CLI is implemented and tested. The dataset is published at EndgameLabs/SalesTranscriptQA on Hugging Face.

Download the pilot and CLI release · Pilot report and costs

The final run accepted 200 of 256 candidates. Median answer length: 13 words. Estimated API spend: $0.89 for the pilot; $0.92 including all development and smoke checks. These are usage-based estimates, not invoices.

The corpus preserves dialogue verbatim. Linked IDs are resolved into separate identifying metadata. Questions may use both dialogue and that metadata; advanced questions must require evidence from exactly two calls within one opportunity or lead.

Publication uses canonical Parquet plus Markdown transcript exports with YAML frontmatter. The salestranscriptqa CLI fetches pinned Hugging Face releases, exports corpora/questions, validates submissions, and supports answer judging in the style of EnronQA-cli.

Attribution and licensing

Source data: Salesforce/CRMArenaPro, accompanying CRMArena-Pro: Holistic Assessment of LLM Agents on Diverse and Realistic Enterprise Tasks, from Salesforce AI Research.

This is an independent adaptation, not an official Salesforce benchmark or a reproduction of its CRM agent evaluation. Original code and documentation in this repository are MIT-licensed. Upstream data and transcript-derived releases retain CC BY-NC 4.0 terms; the MIT license does not relicense them. See NOTICE.md.

About

Dialogue RAG benchmark and CLI for synthetic CRMArena-Pro sales transcripts; specification stage

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages