#readme

Code description

Code for paper : "Who did what? : Creating structured data from acknowledgement texts with large language models".

Use

The code and completions are optimized for the data used in this article. If you want to use it for your data:

1. Add your OpenAI key in the file 'secret_key.py'.
2. Add your own completions in the file 'completions.py'. The code is divided into three sections: keys, few shot learning mapping and instructions. a. Keys tells the model which elements you want to extract with either method: few-shot or instruction. b. Few-show learning requires mapping prompts to completions. In this case, acknowledgements texts were mappped with outputs in JSON format. c. Instruction requires a natural language instruction to run. You can modify these, as long as you keep a similar formatting to the original.

Defaults

Model has a few default values to keep in mind:

1. max_tokens = max tokens to generate (set to 500)
2. stop = sequence where API will stop generating. (None for instructions, '\n\n[Output]:' for few-shot)
3. temperature = level of randomness in the model (set to 0)