Bloom-1b7-creative-writing

This model is a fine-tuned version of bigscience/bloom-1b7 on the adambjorn/UnrelatedForgettingOverhead creative writing dataset.

Model description

More information needed

Intended uses & limitations

Intended for use on a student group project for Portland State University's Winter 2024 LLMs Course.

Training and evaluation data

Instruction Tuned on the creative writing dataset here: https://huggingface.co/datasets/adambjorn/UnrelatedForgettingOverhead/viewer/creative

Training procedure

Trained on a single RTX 3090 card.

Given a set of prompts:

prompts = [
    "Write a creative short story based on the following title:",
    "Here is a title for a story. Craft a short narrative around it:",
    "Using the title given, develop a short story:",
    "Imagine a short story that starts with this title:",
    "Create a brief story with the following title:"
]

Concatenate the prompt, the title and the story like so:

concatenated_texts = [random.choice(prompts) + " " + title + "</s>" + "Story: " + selftext for title, selftext in zip(titles, selftexts)]

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: 5
• mixed_precision_training: Native AMP

Training results

Final results: {'loss': 0.0472, 'learning_rate': 1.4893617021276598e-06, 'epoch': 4.95}

Average results: {'train_runtime': 563.2707, 'train_samples_per_second': 1.687, 'train_steps_per_second': 0.417, 'train_loss': 0.8475136074614018, 'epoch': 4.95}

Framework versions

• Transformers 4.37.2
• Pytorch 2.2.0+cu121
• Datasets 2.17.0
• Tokenizers 0.15.2