experience-model-v1

This model is intended to detect the presence of a present-moment experience a human or animal is experiencing in a sentence.

Usage

Given a sentence, the model gives logits of whether or not that sentence contains a present-moment experience. Higher values correspond to the sentence having that experience.

model = transformers.AutoModelForSequenceClassification.from_pretrained('edmundmills/experience-model-v1')  # type: ignore
tokenizer = transformers.AutoTokenizer.from_pretrained('edmundmills/experience-model-v1', use_fast=False)  # type: ignore
sentence = "I am eating food."
tokenized = tokenizer([sentence], return_tensors='pt', return_attention_mask=True)
input_ids, masks = tokenized['input_ids'], tokenized['attention_mask']
with torch.inference_mode():
    out = model(input_ids, attention_mask=masks)
probs = out.logits.sigmoid().squeeze().item()
print(probs) # 0.92

Model description

This model was fine-tuned from 'microsoft/deberta-v3-large'.

Intended uses & limitations

More information needed

Training and evaluation data

This model was trained on 745 training samples, with ~10% of them containing present moment experiences.

Training procedure

The model was fine-tuned using https://github.com/AlignmentResearch/experience-model. It used BCE Loss.

Training hyperparameters

It used the following hyperparameters: learning_rate: 2.0 e-05 batch_size: 16 epochs: 200 weight_decay: 0.01

Framework versions

• Transformers 4.26.1
• Pytorch 1.13.1
• Datasets 2.9.0
• Tokenizers 0.13.2