codewithyug/alethex-nli-deberta-v3-small

ALETHEX NLI — DeBERTa-v3-Small (Lightweight)

Fast, lightweight contradiction detection — 172MB ONNX, runs in any browser
92.8% accuracy · 92.8% Macro F1 · MCC 0.892 · 22M parameters

---

Overview

This is a fine-tuned version of cross-encoder/nli-deberta-v3-small (22M parameters) for Natural Language Inference (NLI) optimized to detect contradictions between statements made across a conversation.

It is the lightweight variant of the ALETHEX NLI model family. At only 172MB quantized, it loads significantly faster than the large model while achieving the same 92.8% accuracy — ideal for mobile browsers, low-bandwidth connections, or scenarios where first-load time matters.

---

Model Details

Property	Value
Base model	cross-encoder/nli-deberta-v3-small
Architecture	DeBERTa-v2 (small)
Parameters	~22M
Hidden size	384
Task	3-way NLI (contradiction / entailment / neutral)
Training hardware	NVIDIA RTX 4050 Laptop GPU (6.4GB VRAM)
Precision	FP16
Training time	~8 hours
Epochs	3
Batch size	32
Optimizer	AdamW (8-bit, bitsandbytes)
Learning rate	1e-5 with linear warmup (10%)
Weight decay	0.01

---

Performance

Overall Metrics

Metric	Score
Accuracy	92.80%
Macro F1	92.81%
MCC	0.892
Eval Loss	0.2516

Training Progression

Epoch	Accuracy	Macro F1	MCC
1	91.94%	91.94%	0.879
2	92.51%	92.52%	0.887
3	92.80%	92.81%	0.892

---

Labels

ID	Label	Description
0	contradiction	The hypothesis contradicts the premise
1	entailment	The hypothesis follows from the premise
2	neutral	The hypothesis is unrelated to the premise

---

Training Data

Dataset	Pairs	Description
MultiNLI	~393K	Cross-genre NLI
SNLI	~549K	Caption-based NLI
Temporal NLI	~variable	Temporal contradictions
Total	~1M+	Unified, no synthetic data

---

Files

File	Size	Description
model_quantized.onnx	~172 MB	INT8 quantized ONNX — recommended
model.onnx	~568 MB	Full FP32 ONNX
tokenizer.json	<1 MB	Fast tokenizer
config.json	<1 KB	Model config

---

Usage

Transformers.js — In-Browser

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

const clf = await pipeline(
  "zero-shot-classification",
  "codewithyug/alethex-nli-deberta-v3-small",
  { quantized: true }  // loads 172MB model, cached forever
);

const result = await clf(
  "I live in New York.",
  ["contradiction", "same meaning", "unrelated"]
);
console.log(result.labels[0]);  // "contradiction"
console.log(result.scores[0]);  // 0.95

Python — HuggingFace Transformers

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id = "codewithyug/alethex-nli-deberta-v3-small"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

def check_contradiction(premise: str, hypothesis: str) -> dict:
    inputs = tokenizer(premise, hypothesis, return_tensors="pt",
                       truncation=True, max_length=512)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)[0]
    return {model.config.id2label[i]: round(probs[i].item(), 4) for i in range(3)}

print(check_contradiction("I live in New York.", "I moved to London last year."))
# {'contradiction': 0.9521, 'entailment': 0.0239, 'neutral': 0.0240}

Python — Pipeline API

from transformers import pipeline

clf = pipeline("text-classification", model="codewithyug/alethex-nli-deberta-v3-small")
result = clf([{"text": "I live in New York.", "text_pair": "I moved to London."}])
print(result)
# [{'label': 'contradiction', 'score': 0.9521}]

Python — ONNX Runtime

import onnxruntime as ort
from transformers import AutoTokenizer
import numpy as np

tokenizer = AutoTokenizer.from_pretrained("codewithyug/alethex-nli-deberta-v3-small")
session = ort.InferenceSession("model_quantized.onnx")

def infer(premise, hypothesis):
    enc = tokenizer(premise, hypothesis, return_tensors="np",
                    padding="max_length", max_length=128, truncation=True)
    logits = session.run(["logits"], dict(enc))[0]
    probs = np.exp(logits) / np.exp(logits).sum(axis=-1, keepdims=True)
    return {"contradiction": float(probs[0][0]), "entailment": float(probs[0][1]), "neutral": float(probs[0][2])}

---

Small vs Large — Which Should I Use?

	Small (this model)	Large
Parameters	22M	400M
ONNX size	172MB	641MB
First load	~1 min	~3-5 min
Accuracy	92.8%	92.8%
Best for	Mobile, low-bandwidth, fast startup	Desktop, subtle contradictions

---

Example Detections

Earlier statement	Later statement	Label	Confidence
"I live in New York."	"I moved to London last year."	contradiction	0.95
"I'm vegetarian."	"I had steak for dinner."	contradiction	0.92
"I work at Google."	"My job at Microsoft keeps me busy."	contradiction	0.89
"I'm 25 years old."	"I turn 30 next week."	contradiction	0.86

---

Limitations

• English only
• Max 128 tokens per input pair (truncated)
• Not a fact-checker — detects logical contradiction between spans only
• Figurative language / sarcasm may produce false positives

---

Citation

@misc{alethex2026,
  title  = {ALETHEX NLI: Fine-tuned DeBERTa-v3-Small for Temporal Belief Contradiction Detection},
  author = {codewithyug},
  year   = {2026},
  url    = {https://huggingface.co/codewithyug/alethex-nli-deberta-v3-small}
}

---

License

Apache 2.0 — free for commercial and research use.