Skip to content

SAMPLE-DATA.md

Latest commit

 

History

History
69 lines (50 loc) · 2.39 KB

SAMPLE-DATA.md

File metadata and controls

69 lines (50 loc) · 2.39 KB

Sample data

What ships in this repo for benchmarking, what doesn't, and why.

TL;DR

File What it is
training/FROZEN_EVAL_SCORED.jsonl Public eval slice — mixed benign + attack inputs already scored. Lets anyone reproduce our numbers without retraining.
training/CORPUS_LICENSES.json Public benchmark datasets and OSS baseline models we evaluate against, with licenses.
scripts/bench.mjs Re-scores the eval slice against the shipped ONNX. One command.
scripts/bench_oss.py Scores the same inputs against ProtectAI / deepset / fmops. Side-by-side table.

Raw datasets are not redistributed; fetch them from their upstream URLs.

What we deliberately don't ship

  • No raw datasets. Fetch the benchmark sources yourself; see training/CORPUS_LICENSES.json.
  • No teacher weights. Built from scratch.
  • No customer data. No data from any customer system has ever been used for training or evaluation.

How to get the model

Recommended — release tarball + checksum:

TAG=v0.0.1
curl -L -o promptpurify-model.tar.gz \
  https://github.com/securelayer7/PROMPTPurify/releases/download/$TAG/promptpurify-model.tar.gz
curl -L -o promptpurify-model.tar.gz.sha256 \
  https://github.com/securelayer7/PROMPTPurify/releases/download/$TAG/promptpurify-model.tar.gz.sha256
sha256sum -c promptpurify-model.tar.gz.sha256   # MUST print "OK"
tar xzf promptpurify-model.tar.gz                # extracts to ./models/l5e/

The release tarball is cosign-signed with keyless Sigstore; verify with cosign verify-blob --bundle promptpurify-model.tar.gz.cosign.bundle.

Alternative — Hugging Face Hub mirror:

huggingface-cli download Securelayer7/promptpurify --local-dir models/l5e

Alternative — clone the repo (for benchmarking / contributing):

git clone https://github.com/securelayer7/PROMPTPurify
cd promptpurify && shasum -a 256 -c models/l5e/SHA256SUMS

Why the model isn't in the npm tarball

  • Keeps npm install promptpurify tiny (~50 KB).
  • Lets you version the model independently of the SDK.
  • Browser bundle stays ONNX-free — no 14 MB shipped to every visitor.

Reproduce in one command

git clone https://github.com/securelayer7/PROMPTPurify
cd promptpurify
npm install onnxruntime-node
node scripts/bench.mjs

Full step-by-step: REPRODUCE.md.