pszemraj/opt-peter-2.7B

<a href="https://colab.research.google.com/gist/pszemraj/4068382a40bbf7aab50638b062bd97a9/opt-peter-2-7b-example-csearch-generation.ipynb"> <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/> </a>

This model is a fine-tuned version of facebook/opt-2.7b on about 80k WhatsApp/text messages (mine). Please use responsibly :)

Test it out on Google Colab by clicking the button above.

Model description

• Exploring to see how OPT does in terms of dialogue/conversational applications
• Seems to do a lot better than GPT-Neo with similar training parameters
• you can create your own digital clone and deploy it leveraging this repository I am working on.

sharded checkpoint

As this model file is 10+ GB, it can impose some constraints with lower RAM runtimes and/or download speeds. To help with this issue, a sharded checkpoint of this model is available here.

The pszemraj/opt-peter-2.7B-sharded model can be used as a drop-in replacement for this one for all use cases.

Intended uses & limitations

The base model has a custom license that propagates to this one. Most importantly, it cannot be used commercially. Read more here: facebook/opt-2.7b

• the model is probably too large to use via API here. Use in Python with GPU RAM / CPU RAM > 12 GB, Colab notebook linked above.
  • alternatively, you can message a bot on telegram where I test LLMs for dialogue generation
• any statements or claims made by this model do not reflect actual claims/statements by me. Keep in mind it is a fine-tuned version of the model on my data, so things from pre-training are also present in outputs.

Training and evaluation data

WhatsApp & iMessage data were parsed using ai-msgbot and then fed as a text dataset to the HF trainer.

Training procedure

Training hyperparameters

SESSION ONE

The following hyperparameters were used during training:

• learning_rate: 4e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 16
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.01
• num_epochs: 3

SESSION TWO

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 4

Framework versions

• Transformers 4.19.2
• Pytorch 1.10.0+cu113
• Datasets 2.2.2
• Tokenizers 0.12.1