codewithyug/alethex-nli-deberta-v3-large

ALETHEX NLI — DeBERTa-v3-Large

Fine-tuned DeBERTa-v3-Large for 3-class Natural Language Inference (contradiction / entailment / neutral), trained on MultiNLI + SNLI on Nvidia H100 GPU.

Used by the ALETHEX Chrome Extension to detect AI self-contradictions in real time, running 100% in-browser via Transformers.js.

Model Details

Property	Value
Base model	microsoft/deberta-v3-large
Task	3-class NLI (contradiction / entailment / neutral)
Training data	MultiNLI + SNLI (~942K pairs)
Accuracy	92.8% on MultiNLI matched dev
Training hardware	Nvidia H100 80GB
ONNX FP32 size	1,742 MB
ONNX INT8 size	641 MB (2.71× compression)

Labels

id2label = {0: "contradiction", 1: "entailment", 2: "neutral"}

Usage — PyTorch

from transformers import pipeline

nli = pipeline("zero-shot-classification", model="codewithyug/alethex-nli-deberta-v3-large")

result = nli(
    "I live in New York.",
    candidate_labels=["contradiction", "entailment", "neutral"],
    hypothesis_template="This statement {}s the original: I moved to London last year."
)
print(result)

Usage — ONNX via Transformers.js (browser)

import { pipeline } from "@xenova/transformers";

// quantized=true loads onnx/model_quantized.onnx (~641MB, INT8)
const clf = await pipeline(
  "zero-shot-classification",
  "codewithyug/alethex-nli-deberta-v3-large",
  { quantized: true }
);

const out = await clf(
  "I am vegetarian.",
  ["contradiction", "entailment", "neutral"],
);
console.log(out);
// { labels: ["contradiction", ...], scores: [0.94, ...] }

ONNX Files

File	Size	Description
onnx/model.onnx	1,742 MB	FP32 — full precision, best accuracy
onnx/model_quantized.onnx	641 MB	INT8 dynamic quantization — used by the Chrome extension

Training

Trained with Hugging Face Trainer on MultiNLI + SNLI combined:

• 3 epochs, batch size 32, learning rate 2e-5
• Warm-up ratio 0.06, weight decay 0.01
• FP16 mixed precision on H100

ALETHEX Project

This model powers the ALETHEX contradiction detection engine:

• Chrome Extension — real-time AI consistency monitoring on ChatGPT, Claude, Gemini, Copilot, DeepSeek, Perplexity, Grok
• Python Library — pip install alethex for RAG pipelines and LangChain/LlamaIndex integrations

License

Apache 2.0