Animised/nli-cdv3

Animised NLI Contradiction Detector v3

prajjwal1/bert-medium (41M) trained directly on hard labels
with a 3:1 imbalanced dataset to prevent contradiction bias.

Why v3?

Upgrade from bert-small to bert-medium for stronger performance,
while keeping the conservative contradiction policy from v2.

Results

Metric	Value
Accuracy	0.8636 (86.36%)
Loss	0.366860
Epochs	4

Labels

0 = entailment | 1 = neutral | 2 = contradiction

Usage

from transformers import pipeline
clf = pipeline("text-classification", model="Animised/nli-cdv3")
clf("Premise [SEP] Hypothesis", top_k=None)

Purpose

Character fact consistency checker for the
Animised project.

Training details

• Base model : prajjwal1/bert-medium (41M params)
• Dataset : Animised/nli-v3
• Data ratio : 3:1 (entailment+neutral : contradiction)
• Loss : CrossEntropyLoss (hard labels)
• Epochs : 4
• Batch size : 384
• Max length : 256
• LR : 4e-05
• GPUs : 2