kjnlae DistilBERT v0.2 — Experimental Prompt-Injection Classifier

Experimental research artifact. NOT production-ready security software. This model can miss malicious instructions or flag benign text. Never rely on it as the sole safeguard for tools, resources, secrets or regulated systems. Model scores are NOT calibrated probabilities of attack success.

Fine-tuned locally from Apache-2.0 licensed DistilBERT base uncased, pinned revision 12040accade4e8a0f71eabdb258fecc2e7e948be.

The weights implement the sole prompt-injection detector in kjnlae, an offline-first Python security gateway with separate application-controlled authorization and output-release checks.

Labels: NON_INJECTION (0) and INJECTION (1). English only.

Verified model manifest SHA-256: 54e6df56616f4533e1c6a9624f3372ad14bc4d85f5ae97b45c1b3b515a83de9b.

Intended use and limitations

Experimental local scanning of untrusted messages, email, webpages, retrieval chunks and tool responses. Pair the classifier with host-verified principal, tenant, grants, tool allowlists, human approvals, resource versions and output release controls. A negative scan never grants action permission. Not validated for production, healthcare, regulated records, or deployment against real autonomous attacks.

Training provenance

The model adapts the pretrained English DistilBERT encoder and a newly trained two-class classifier head. The final training set had 924 rows (639 benign, 285 injection). Separate partitions had 12 validation, 12 calibration and 116 Deepset held-out rows. Sources:

• Original fictional kjnlae training cases and hard benign examples (provenance).
• Microsoft BIPIA active attack prompts, pinned revision a004b69ec0dd446e0afd461d98cb5e96e120a5d0 (retain LICENSE and NOTICE).
• Deepset prompt-injections, pinned revision 4f61ecb038e9c3fb77e21034b22511b523772cdd, plus TRAIN-only source-balanced benign variants. The dataset's top-level metadata reports Apache-2.0 but nested metadata reports CC BY 4.0; see source notices and the license/provenance discussion.

No raw Deepset or BIPIA examples are included in this model repository.

Training used four CPU cores, batch size 2, 128 tokens, AdamW learning rate 2e-5, seed 42, and two epochs (924 total optimizer steps), taking 351.13 seconds. Validation selected epoch 1 as the best checkpoint, so the distributed weights represent 462 optimization steps, not the final epoch.

The gateway's calibration threshold is 0.025134501978754997, chosen using only 12 calibration rows. This very low value is unstable and likely contributes to false positives. A standard Hugging Face text-classification pipeline does not implement kjnlae's text preprocessing, complete-window coverage, calibration threshold or authorization security boundaries, and should not be treated as an equivalent prompt-injection defense.

Measured evaluation of the reviewed weights

The sample sizes are small and evaluation distributions differ. The earlier model achieved 0% false positives on the Deepset split, so this follow-up was a regression there even though the independent 50-case fixture improved. The current model frequently overflags code and educational quotations. These are not general deployment accuracy estimates.

The separately controlled gateway completed 2/2 authorized tasks and blocked 4/4 unauthorized actions in a six-case scripted synthetic sandbox, including one classifier miss. There was no live LLM. AgentDojo and full external BIPIA benchmarks were not run.

Full evaluation report · Benchmark methodology · Known limitations.

Install and run locally

Clone and install kjnlae with the CPU-only dependencies described in its README, and install huggingface_hub>=0.30,<1. Then, from the project root:

kjnlae model download --repo namelessweakling/kjnlae-distilbert-v0.2
export KJNLAE_MODEL_DIR="$PWD/var/models/full-v0.2-v3-reviewed"
export KJNLAE_MODEL_HASH="54e6df56616f4533e1c6a9624f3372ad14bc4d85f5ae97b45c1b3b515a83de9b"
kjnlae model verify "$KJNLAE_MODEL_DIR" --hash "$KJNLAE_MODEL_HASH"
kjnlae doctor
kjnlae serve

The downloader installs the files only after comparing the manifest and each file with the pinned hashes. The normal scanning path makes no network requests. The destination must not exist before download.

License and attribution

Proposed release license for the fine-tuned weights: Apache-2.0 (the base encoder is Apache-2.0). The kjnlae application code is separately MIT licensed. Distribution of these weights does not override any obligations from the Microsoft BIPIA or Deepset training sources. Refer to their licenses, notices, and the linked provenance report; the Deepset metadata conflict warrants extra review before downstream redistribution or commercial use.

Maintainer: confirm the Apache-2.0 model licensing and attribution before uploading publicly.