ettin-nli-classifier

Ettin-NLI-Classifier

This model uses the jhu-clsp/ettin-encoder-68m architecture, which has been fine-tuned for NLI tasks on the MultiNLI and SNLI datasets. It was made to replace the old distilroberta finetune. It achieves the following results on the evaluation set:

• Accuracy: 0.88142
• F1: 0.88158
• Precision: 0.88182
• Recall: 0.88142

Model description

The SNLI corpus (version 1.0) is a collection of 570k human-written English sentence pairs manually labeled for balanced classification with the labels entailment, contradiction, and neutral, supporting the task of natural language inference (NLI), also known as recognizing textual entailment (RTE).

The Multi-Genre Natural Language Inference (MultiNLI) corpus is a crowd-sourced collection of 433k sentence pairs annotated with textual entailment information. The corpus is modeled on the SNLI corpus, but differs in that covers a range of genres of spoken and written text, and supports a distinctive cross-genre generalization evaluation.

Usage

Inference API has been disabled as it is not suitable for this kind of task.

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_checkpoint = 'AdamCodd/ettin-nli-classifier'
model = AutoModelForSequenceClassification.from_pretrained(model_checkpoint)
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)

# Set device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# Sample premise and hypothesis
premise = "The cat is sleeping under the sun."
hypothesis = "It's raining, and the cat is getting wet."

# Tokenize and predict
input = tokenizer(premise, hypothesis, truncation=True, padding=True, return_tensors="pt", max_length=256).to(device)
with torch.no_grad():
    output = model(**input)
    probabilities = torch.softmax(output.logits, dim=-1)[0].tolist()

# Output prediction
label_names = ["Entailment", "Neutral", "Contradiction"]
prediction = {name: round(prob * 100, 1) for name, prob in zip(label_names, probabilities)}
print(prediction)
# e.g. {'Entailment': 1.3, 'Neutral': 8.2, 'Contradiction': 90.5}

Training and evaluation data

The training data consists of a concatenated corpus of the SNLI train split and the MultiNLI train split. The evaluation metrics were calculated using a concatenated validation set consisting of the SNLI validation split and the MultiNLI validation_matched split. All -1 labels (samples without annotator consensus) were filtered out prior to training.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 32
• eval_batch_size: 64
• seed: 42
• optimizer: AdamW with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• warmup_ratio: 0.05
• num_epochs: 1
• weight_decay: 0.01
• mixed_precision: fp16

Training results

Metrics: Accuracy, F1, Precision, Recall

'eval_loss': 0.62099,
'eval_accuracy': 0.88142,
'eval_f1': 0.88158,
'eval_precision': 0.88182,
'eval_recall': 0.88142

NB: The confusion matrix is here.

Framework versions

• Transformers >= 4.48.0
• Datasets >= 2.16.1
• Evaluate >= 0.4.1
• Tokenizers >= 0.15.0

If you want to support me, you can here.