SriRamanaAtmic/AtmicNLI

Fine-tuned from MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli for natural language inference (entailment / neutral / contradiction) on Sri Ramana Maharshi teaching-corpus query/span pairs.

Training data

• 723 train examples, 183 held-out validation examples
• Validation split is grouped by original query (paraphrases of the same query are never split across train/val) and stratified by label

Validation results

               precision    recall  f1-score   support

   entailment     0.6081    0.6522    0.6294        69
      neutral     0.7907    0.6296    0.7010        54
contradiction     0.3939    0.4333    0.4127        60

     accuracy                         0.5738       183
    macro avg     0.5976    0.5717    0.5810       183
 weighted avg     0.5918    0.5738    0.5795       183

Confusion matrix (rows=gold, cols=predicted):
                  entailment       neutral contradiction
    entailment            45             3            21
       neutral             1            34            19
 contradiction            28             6            26

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("SriRamanaAtmic/AtmicNLI")
model = AutoModelForSequenceClassification.from_pretrained("SriRamanaAtmic/AtmicNLI")

premise = "..."   # span / context
hypothesis = "..."  # query / claim
inputs = tokenizer(premise, hypothesis, truncation="only_first", return_tensors="pt")
logits = model(**inputs).logits