distilgpt2-emailgen

Why write the rest of your email when you can generate it?

from transformers import pipeline

model_tag = "postbot/distilgpt2-emailgen"
generator = pipeline(
              'text-generation', 
              model=model_tag, 
            )
            
prompt = """
Hello, 

Following up on the bubblegum shipment."""

result = generator(
    prompt,
    max_length=64,
    do_sample=False,
    early_stopping=True,
) # generate
print(result[0]['generated_text'])

• try it in a Google Colab notebook
• Use it in bash/cmd with this gist :)

For this model, formatting matters. The results may be (significantly) different between the structure outlined above and prompt = "Hey, just wanted to ..." etc.

Model description

This model is a fine-tuned version of distilgpt2 on a dataset of 50k emails, including the classic aeslc dataset.

It achieves the following results on the evaluation set:

• Loss: 2.6247

Intended uses & limitations

The intended use of this model is to provide suggestions to "autocomplete" the rest of your email. Said another way, it should serve as a tool to write predictable emails faster. It is not intended to write entire emails; at least some input is required to guide the direction of the model.

Please verify any suggestions by the model for A) False claims and B) negation statements before accepting/sending something.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 32
• total_train_batch_size: 256
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.02
• num_epochs: 5

Training results

Training Loss	Epoch	Step	Validation Loss
2.8299	1.0	248	2.7971
2.6984	2.0	496	2.6826
2.7022	3.0	744	2.6361
2.6436	4.0	992	2.6245
2.6195	5.0	1240	2.6247

Framework versions

• Transformers 4.21.1
• Pytorch 1.12.0+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	24.89
ARC (25-shot)	21.76
HellaSwag (10-shot)	27.52
MMLU (5-shot)	25.97
TruthfulQA (0-shot)	46.17
Winogrande (5-shot)	51.62
GSM8K (5-shot)	0.0
DROP (3-shot)	1.16