Modelpublic

enguard/tiny-guard-8m-en-prompt-harmfulness-binary-mix

sourceHugging Facemitupdated 11mo agoView on Hugging Face

0likes7downloads

Model Card

enguard/tiny-guard-8m-en-prompt-harmfulness-binary-mix

This model is a fine-tuned Model2Vec classifier based on minishlab/potion-base-8m for the prompt-harmfulness-binary found in the nicholasKluge/harmful-text dataset.

Installation

bash

pip install model2vec[inference]

Usage

python

from model2vec.inference import StaticModelPipeline

model = StaticModelPipeline.from_pretrained(
  "enguard/tiny-guard-8m-en-prompt-harmfulness-binary-mix"
)


# Supports single texts. Format input as a single text:
text = "Example sentence"

model.predict([text])
model.predict_proba([text])

Why should you use these models?

—Optimized for precision to reduce false positives.
—Extremely fast inference: up to x500 faster than SetFit.

This model variant

Below is a quick overview of the model variant and core metrics.

Field	Value
Classifies	prompt-harmfulness-binary
Base Model	minishlab/potion-base-8m
Precision	0.9522
Recall	0.9100
F1	0.9306

Confusion Matrix

True \ Predicted	FAIL	PASS
FAIL	3813	356
PASS	176	4024

<details> <summary>Full metrics (JSON)</summary>

json

{
  "FAIL": {
    "precision": 0.9522128699429813,
    "recall": 0.9099636741048261,
    "f1-score": 0.9306089956216002,
    "support": 3854.0
  },
  "PASS": {
    "precision": 0.9205222171323866,
    "recall": 0.9580452920143028,
    "f1-score": 0.938909005957248,
    "support": 4195.0
  },
  "accuracy": 0.9350229842216424,
  "macro avg": {
    "precision": 0.9363675435376839,
    "recall": 0.9340044830595644,
    "f1-score": 0.9347590007894241,
    "support": 8049.0
  },
  "weighted avg": {
    "precision": 0.9356962481837012,
    "recall": 0.9350229842216424,
    "f1-score": 0.9349348178800226,
    "support": 8049.0
  }
}

</details>

<details> <summary>Sample Predictions</summary>

Text	True Label	Predicted Label
answer explain reviewing particular piece kryptonite	PASS	PASS
also remove old talk page discussions page discretion need speak please post new messages bottom talk page sign four tildes left note talk page likely respond comments either incivil flame bait personal attacks trolling deleted signpost deliveries automatically archived user talk sjones signpost thanks	PASS	PASS
A new task force to tackle mutilation was being launched, the panel said.	PASS	PASS
much quoting barnes lipstadt used source quote barnes words much need sources either put work context historians journalists responding time finally done near end respond points historians written recently many deaths destruction allied bombing human cost expulsion ethnic germans eastern europe drawing conclusions acknowledging abuses repeated according quotes article quotes thing said example add anything thought enough assessment others thought	PASS	PASS
angle sucks find better one	FAIL	FAIL
answer explain reviewing particular piece kryptonite	PASS	PASS

</details>

<details> <summary>Prediction Speed Benchmarks</summary>

Dataset Size	Time (seconds)	Predictions/Second
1	0.0002	6364.65
1000	0.0784	12752.99
8369	0.6279	13329.32

</details>

Other model variants

Below is a general overview of the best-performing models for each dataset variant.

Classifies	Model	Precision	Recall	F1
prompt-harmfulness-binary	enguard/tiny-guard-2m-en-prompt-harmfulness-binary-mix	0.9533	0.8640	0.9065
prompt-harmfulness-binary	enguard/tiny-guard-4m-en-prompt-harmfulness-binary-mix	0.9570	0.8941	0.9245
prompt-harmfulness-binary	enguard/tiny-guard-8m-en-prompt-harmfulness-binary-mix	0.9522	0.9100	0.9306
prompt-harmfulness-binary	enguard/small-guard-32m-en-prompt-harmfulness-binary-mix	0.9579	0.9141	0.9355
prompt-harmfulness-binary	enguard/medium-guard-128m-xx-prompt-harmfulness-binary-mix	0.9558	0.8923	0.9230

Resources

—Awesome AI Guardrails: <https://github.com/enguard-ai/awesome-ai-guardails>
—Model2Vec: https://github.com/MinishLab/model2vec
—Docs: https://minish.ai/packages/model2vec/introduction

Citation

If you use this model, please cite Model2Vec:

@software{minishlab2024model2vec,
  author       = {Stephan Tulkens and {van Dongen}, Thomas},
  title        = {Model2Vec: Fast State-of-the-Art Static Embeddings},
  year         = {2024},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.17270888},
  url          = {https://github.com/MinishLab/model2vec},
  license      = {MIT}
}