snowflake-arctic-embed-s-nli

• Base Model: Snowflake/snowflake-arctic-embed-s
• Task: Natural Language Inference (NLI)
• Framework: Hugging Face Transformers, Sentence Transformers

snowflake-arctic-embed-s-nli is a fine-tuned NLI model that classifies the relationship between pairs of sentences into three categories: entailment, neutral, and contradiction. It enhances the capabilities of Snowflake/snowflake-arctic-embed-s for improved performance on NLI tasks.

Intended Use

snowflake-arctic-embed-s-nli is ideal for applications requiring understanding of logical relationships between sentences, including:

• Semantic textual similarity
• Question answering
• Dialogue systems
• Content moderation

Performance

snowflake-arctic-embed-s-nli was trained on the sentence-transformers/all-nli dataset, achieving competitive results in sentence pair classification.

Performance on the MNLI matched validation set:

• Accuracy: 0.7745
• Precision: 0.78
• Recall: 0.77
• F1-score: 0.77

Training details

Training Details

• Dataset:

  • Used sentence-transformers/all-nli.

• Sampling:

  • 100 000 training samples and 10 000 evaluation samples.

• Fine-tuning Process:

  • Custom Python script with adaptive precision training (bfloat16).
  • Early stopping based on evaluation loss.

• Hyperparameters:

  • Learning Rate: 2e-5
  • Batch Size: 64
  • Optimizer: AdamW (weight decay: 0.01)
  • Training Duration: Up to 10 epochs

Reproducibility

To ensure reproducibility:

• Fixed random seed: 42
• Environment:
  • Python: 3.10.12
  • PyTorch: 2.5.1
  • Transformers: 4.44.2

Usage Instructions

Using Sentence Transformers

from sentence_transformers import CrossEncoder

model_name = "agentlans/snowflake-arctic-embed-s-nli"
model = CrossEncoder(model_name)
scores = model.predict(
    [
        ("A man is eating pizza", "A man eats something"),
        (
            "A black race car starts up in front of a crowd of people.",
            "A man is driving down a lonely road.",
        ),
    ]
)

label_mapping = ["entailment", "neutral", "contradiction"]
labels = [label_mapping[score_max] for score_max in scores.argmax(axis=1)]
print(labels)
# Output: ['entailment', 'contradiction']

Using Transformers Library

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "agentlans/snowflake-arctic-embed-s-nli"
model = AutoModelForSequenceClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

features = tokenizer(
    [
        "A man is eating pizza",
        "A black race car starts up in front of a crowd of people.",
    ],
    ["A man eats something", "A man is driving down a lonely road."],
    padding=True,
    truncation=True,
    return_tensors="pt",
)

model.eval()
with torch.no_grad():
    scores = model(**features).logits
    label_mapping = ["entailment", "neutral", "contradiction"]
    labels = [label_mapping[score_max] for score_max in scores.argmax(dim=1)]
    print(labels)
# Output: ['entailment', 'contradiction']

Limitations and Ethical Considerations

snowflake-arctic-embed-s-nli may reflect biases present in the training data. Users should evaluate its performance in specific contexts to ensure fairness and accuracy.

Conclusion

snowflake-arctic-embed-s-nli offers a robust solution for NLI tasks, enhancing Snowflake/snowflake-arctic-embed-s's capabilities with straightforward integration into existing frameworks. It aids developers in building intelligent applications that require nuanced language understanding.