Bloom-1b7-dialogsum-IT-baseline

Bloom-1b7-dialogsum-IT

This model is a instruction-tuned version of bigscience/bloom-1b7 on a dialog summation dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

Instruction Tuned on the dialog summation task here: https://huggingface.co/datasets/adambjorn/UnrelatedForgettingOverhead/viewer/dialogsum/train

Training procedure

Given a set of prompts:

prompts = [
    "Provide a concise summary for the following dialogue:",
    "Summarize this conversation in a few sentences:",
    "Here is a dialogue. Can you summarize it briefly?",
    "Read the following dialogue and write a short summary:",
    "Condense the essence of this conversation into a summary:"
]

Each example is concatenated with the prompt, the dialogue, and the summary as so:

    concatenated_texts = [
        random.choice(prompts) + " " + dialogue + "<\s>" + " Summary:" + summary
        for dialogue, summary in zip(examples['dialogue'], examples['summary'])
    ]

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Final epoch results: {'loss': 0.0137, 'grad_norm': 0.6599154472351074, 'learning_rate': 7.000000000000001e-07, 'epoch': 10.0}

Average results: {'train_runtime': 1142.1524, 'train_samples_per_second': 1.751, 'train_steps_per_second': 0.438, 'train_loss': 0.37129621666669843, 'epoch': 10.0}

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.0+cu121
• Datasets 2.17.0
• Tokenizers 0.15.2