README.md

metadata

license: apache-2.0
tags:
  - generated_from_trainer
  - text-generation
  - opt
  - non-commercial
  - dialogue
  - chatbot
widget:
  - text: 'If you could live anywhere, where would it be? peter szemraj:'
    example_title: live anywhere
  - text: 'What would you sing at Karaoke night? peter szemraj:'
    example_title: Karaoke
  - text: >-
      If you could hire someone to help you, would it be with cleaning, cooking,
      or yard work? peter szemraj:
    example_title: help
  - text: >-
      What form of public transportation do you prefer? (air, boat, train, bus,
      car, etc.) peter szemraj:
    example_title: transportation
  - text: 'What''s your favorite zoo animal? peter szemraj:'
    example_title: animal
  - text: 'Do you like or dislike surprises? Why or why not? peter szemraj:'
    example_title: surprises
  - text: >-
      What celebrity would you like to meet at Starbucks for a cup of coffee?
      peter szemraj:
    example_title: 'celebrity '
inference:
  parameters:
    min_length: 2
    max_length: 64
    length_penalty: 0.7
    temperature: 0.65
    no_repeat_ngram_size: 2
    top_k: 20
    do_sample: true
    repetition_penalty: 4.5

pszemraj/opt-peter-1.3B

This model is a fine-tuned version of pszemraj/opt-peter-1.3B-1E on 80k Whatsapp/iMessages (mine).

It achieves the following results on the evaluation set, after training for 1 epoch (on top of the 1E checkpoint linked above):

• eval_loss: 3.4220
• eval_runtime: 954.9678
• eval_samples_per_second: 9.114
• eval_steps_per_second: 2.279
• epoch: 1.0
• step: 1235

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

Intended uses & limitations

• OPT has a license that does not allow for commercial use, see original for details
• any statements or claims made by this model do not reflect actual claims/statements by me

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 16
• 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: 0.01
• num_epochs: 2

Framework versions

• Transformers 4.19.2
• Pytorch 1.11.0+cu113
• Tokenizers 0.12.1