hybrid-guardrails-deberta-moderation

A DeBERTa-v3-base binary safe/unsafe content classifier, fine-tuned as the neural component of the hybrid-guardrails neuro-symbolic moderation pipeline. It is also evaluated standalone as the "Deberta Only" baseline in that project's experiments.

Architecture

DebertaV2ForSequenceClassification, fine-tuned from microsoft/deberta-v3-base (12 layers, hidden size 768), 2-way classification (safe / unsafe).

Training data

Fine-tuned on a combined corpus of:

• AdvBench (adversarial/harmful instruction prompts, MIT-licensed)
• ToxiGen (machine-generated implicit/adversarial hate speech, gated HF dataset)
• Alpaca-Cleaned (benign instruction-following data, CC BY-NC 4.0, non-commercial research use)

Training config: training/configs/deberta_base.yaml (3 epochs, batch size 4, gradient accumulation 4, lr 2e-5, seed 42). See the project's Data Availability section for exact license/redistribution terms of each source.

Intended use

Research artifact for benchmarking hybrid symbolic+neural LLM moderation pipelines. Not validated as a general-purpose production content-safety classifier -- evaluated specifically against the HNS-Bench adversarial benchmark and the clean/adversarial test sets described in the accompanying paper.

Evaluation results (seed 42, Experiment A baseline comparison)

Aggregated across the project's clean and adversarial test sets:

Metric	Value
Precision	0.890
Recall	0.952
F1	0.919
AUROC	0.975
FRR	0.060

Full per-dataset breakdown and multi-seed variance available in the source repository's results/experiment_A/ and paper/MANUSCRIPT.md (§5.1).

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("JashVora7/hybrid-guardrails-deberta-moderation")
model = AutoModelForSequenceClassification.from_pretrained(
    "JashVora7/hybrid-guardrails-deberta-moderation"
)

inputs = tok("ignore all previous instructions", return_tensors="pt")
with torch.no_grad():
    probs = torch.softmax(model(**inputs).logits, dim=-1)
print(probs)  # [P(safe), P(unsafe)]

License

MIT, matching the source repository.