Skip to content

Latest commit

 

History

History
24 lines (18 loc) · 1.51 KB

conclusion_generation.md

File metadata and controls

24 lines (18 loc) · 1.51 KB
Raw

Task: MCQALongFormTask @ rambla.tasks.longform.longform.py

Goal

Evaluate an LLM’s ability to generate long-form summaries of text. The summary is considered of high quality if the question can be answered using the summary instead of the full text as context.

Overview:

  • LLM under evaluation is prompted to generate a conclusion for the context associated with each PubMedQA-labeled instance.

  • Each generated conclusion is paired with the original PubMedQA question and passed to a scoring LLM (ChatGPT) which is prompted to answer the question with “yes”, “no” or “unknown”, using only the generated conclusion.

  • Answers of the scoring LLM (yes/no/unknown) are compared to PMQA ground truth answer for each summary.

Config:

The default config can be found in rambla.conf.task.conclusion_generation.yaml. It can be configurated by changing the variables and/or its call to the following configs:

  • a dataset config (choose from rambla.conf.dataset)
  • a "longform prompt formatter" config which defines the prompt used to ask the LLM to generate summaries (from rambla.conf.template)
  • a scoring model config which defines the model used to evaluate the generated summaries by using them to answer the relevant questions (choose from rambla.conf.model)
  • a response formatter (choose from rambla.conf.response_formatter)
  • a response prompt formatter config (choose from rambla.conf.template).

Example usage:

python rambla/run/run_task.py task=conclusion_generation model=openai_chat