Gatekeeper-Scope

Training and evaluation data for scope enforcement ("is this prompt something my app is meant to handle?") and prompt routing ("which handler should get it?"), used to train Gatekeeper.

config / split	rows
scope / train	35,966
scope / validation	3,000
scope / test	5,000
route / train	21,984
route / validation	1,000
route / test	3,500

• scope: text, purpose (a one-sentence app description), in_scope (0/1), profile, pool.
• route: text, routes (JSON name -> description), with_other, target, pool.
• Held-out profiles (medical_info, sql_helper) appear only in test, to measure zero-shot scope enforcement.
• About 15% of out-of-scope negatives are attack prompts aimed at the app.

Prompts come from public datasets; each keeps its original license, so check the source before commercial use:

• Gustavosta/Stable-Diffusion-Prompts
• sahil2801/CodeAlpaca-20k
• openai/gsm8k
• bitext/Bitext-customer-support-llm-chatbot-training-dataset
• keivalya/MedQuad-MedicalQnADataset
• b-mc2/sql-create-context
• databricks/databricks-dolly-15k
• clinc/clinc_oos
• nuhmanpk/prompt-sentinel (attack negatives)

Built by Nuhman PK.