distillbert_conv_quality_score

This model is a fine-tuned version of distilbert-base-uncased on the conv_ai_2 dataset. It was trained to generate a score (in the [0, 1] range) from a conversation

It achieves the following results on the evaluation set:

• training/loss: 0.0165
• validation/loss: 0.0149

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification

model_name = "alespalla/distillbert_conv_quality_score"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

conversation = '''
Q: Begin
A: lol ! do you think it is strange to feel like you have been through life before ?
Q: Hellow
A: I don't understand you 🙈. Also, try to guess: i like to ...
Q: How are you?
A: make time stop, funny you :)
Q: What is your name?
A: jessie. hows your day going ? 😃
'''

score = model(**tokenizer(conversation, return_tensors='pt')).logits.item()
print(f"Score: {score}")

Training and evaluation data

The training data was generated from conv_ai_2 using the following function

from datasets import load_dataset

def get_dataset(regression=False):

    db = load_dataset("conv_ai_2")

    def generate_converation(elem):
        text = ""
        for idx, txt in enumerate(elem["dialog"]):
            if idx % 2:
                text += f"A: {txt['text']}\n"
            else:
                text += f"Q: {txt['text']}\n"
        if regression:
            return {'text': text, "labels": (elem['eval_score'] - 1)/4}
        return {'text': text, "labels": elem['eval_score'] - 1}

    db = db.filter(lambda example: example["eval_score"] > 0)
    db = db.map(generate_converation, remove_columns=db['train'].column_names)
    db = db['train'].train_test_split(test_size=0.2).shuffle(42)

    return db

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• epochs: 40
• batch_size: 16
• learning_rate: 0.0002
• eval_steps: 82
• log_steps: 82
• save_steps: 41
• gradient_accumulation_steps: 1
• warmup_steps: 0

Training results

step	training/loss	validation/loss
81	0.1020	0.0794
163	0.0800	0.0713
245	0.0553	0.0491
327	0.0362	0.0440
409	0.0282	0.0352
491	0.0282	0.0412
573	0.0256	0.0293
655	0.0238	0.0252
737	0.0175	0.0226
819	0.0154	0.0228
901	0.0116	0.0205
983	0.0160	0.0202
1065	0.0146	0.0240
1147	0.0182	0.0180
1229	0.0171	0.0192
1311	0.0091	0.0174
1393	0.0171	0.0158
1475	0.0137	0.0158
1557	0.0158	0.0148
1639	0.0165	0.0149

Framework versions

• Transformers 4.26.1
• Datasets 2.10.1
• Tokenizers 0.13.2