ai4labs/gpt3playaround

GPT-3 Experiments for Sentiment and Arousal Analysis

##Instructions:

• Create a clean Anaconda virtual environment (using Python 3.7) through Anaconda Navigator (this avoids unnecessary package installations through Anaconda Prompt)

• Install required packages: pip install -r requirements_gpt3.txt

• Specify all necessary input configs in input_config.json.

• If you're labelling Sentiment, it is recommended that you specify "train_mode/train_or_zeroshot" as "train" to prime the model with training instances, specify "prompt_type" under "prompt_design" in the config as "long", use engine "davinci-instruct-beta" or "cushman-alpha" for higher accuracy, replace "prompt_design/prompt" with your prompt of choice, choose "train" under "train_or_zeroshot", train with 1 instance/category in "num_lines_per_category" and choose batch size 5 in "num_lines_per_batch". An ideal prompt in this situation would look like: (also format of "long" prompt)

  

Then specify input prefix (replace "Tweet") and output prefix ("Output") of choice, specify number of training instances per category you want to prime GPT-3 with (to maintain balance in training data). Note that input text file should be in the format

• If you're labelling Arousal, it is recommended that you specify "train_mode/train_or_zeroshot" as "zeroshot" to perform zeroshot training, specify "prompt_type" under "prompt_desisgn" in the config as "short", use engine "davinci-instruct-beta" for higher accuracy, replace "prompt_design/prompt" with your own descriptive prompt. An ideal prompt in this situation would look like: (also format of "short" prompt)

Explanations of config variables:

• gpt3_engine: OpenAI's GPT-3 engine to that takes in prompt inputs and generates output text. All config setting variables in this section pertain to OpenAI's trainable parameters as documented through OpenAI's official documentation.

• classification_task: Choose between "valence" and "arousal" for Sentiment and Arousal Classification, respectively.

• train_or_zeroshot: Choose between passing in training instances to prime the transformer ("train") or doing zeroshot learning ("zeroshot")(using pretrained weights to make classification predictions)

• output_dir: Directory to put prediction output text file into.

• output_filename: Desired name of output file.

• num_lines_per_category: If "train_or_zeroshot" is specified as "train", the number of training instances per sentiment/arousal category to prime the transformer with.

• num_lines_per_batch: If "prompt_type" is specified as "long", the number of instances to be passed to the transformer to get GPT-3 output from.

• input_file_to_label: Testing input file whose sentences will be passed to GPT-3 to obtain classification output for. Note: This file should be formatted with no heading, with each sentence being in the format: "Sentiment/Arousal_score, Phrase_text\n"

• train_input_file: If "train_or_zeroshot" is specified as "train", the input file to obtain training instances from. Note: This file should be formatted with heading "Sentiment_class_label:->Phrase_text", with each sentence having the format "Sentiment/Arousal_score:->Phrase_text\n"

• descriptive_prompt: If "train_or_zeroshot" is specified as "zeroshot", the prompt to use as a single sentence to prime the transformer before testing input is sent.

• prompt: If "train_or_zeroshot" is specified as "train", the prompt to use to prime the transformer before testing input is sent.

• prompt_type: "long" or "short" (see example above)

• input_prefix: Prefix to go before passing input sentences

• output_prefix: Prefix to nudge GPT-3 to return classification prediction.