Use Llama-2 for Effective Text Summarization

Contributed by: Federico Castanedo

Difficulty level: Intermediate

Text summarization is the process of generating concise and coherent representations of longer texts while keeping the most essential information. The objective is to capture the relevant information in a shortened form.

Text summarization is a complex and challenging task for several reasons. The summary should be representative of the entire text and not biased toward a particular segment. It is also important to avoid repetition while ensuring all the main concepts are covered.

Text summarization has been a common use case for LLMs and chatbots, but users usually struggle to find good summaries from LLM output.

Recently, a team of researchers from Salesforce AI, MIT, and Columbia presented a technique named a “Chain of Density” (CoD) prompt to generate automatic summaries. The summaries they obtain with the proposed method are more abstractive, exhibit more fusion, and are less biased than the ones generated by a standard vanilla prompt.

Inspired by the CoD prompt technique proposed in “From Sparse to Dense: GPT-4 summarization with Chain of Density Prompting,” I used Llama-2-70b-chat-hf from HuggingChat with a slightly revised version of the paper’s prompt template to yield better results. You can insert your text in the ARTICLE section of the following template to have Llama summarize it:

Prompt: Article: {{ ARTICLE }}

Please generate increasingly concise, ...