Various complex tasks can be easily solved using Language Models and Retrieval models. Language models, like GPT-3, are designed to generate human-like text based on the input they receive. On the other hand, retrieval models are used to retrieve relevant information from a database or collection of documents. Clearly defining the task you want to solve determines whether it requires generating new text or retrieving data from existing sources.
With GPT-3 or similar models, one needs to provide a prompt describing the task and let the model generate text based on it. It requires experimenting with the prompt wording and structure to get the desired output. It involves combining generated text from language models with information retrieved from databases. This could include generating summaries or insights based on retrieved information.
Researchers at Stanford build a framework for solving advanced tasks with language models(LMs) and retrieval models(RMs). They call it as DSPy. DSPy consists of various techniques for prompting and finetuning the LMs and improving their reasoning and retrieval augmentation. DSPy is based on Pythonic syntax to provide composable and declarative modules for instructing LMs.
DSPy also has an automatic compiler that trains the LM to run the declarative steps in your program. This compiler can finetune finetune from minimal data without manual intermediate-stage labels. It uses systematic space of modular and trainable pieces instead of string manipulation.
DSPy uses two simple concepts called ?Signatures? and ?Teleprompters? to compile any program you write. A signature is a declarative specification of the input/output behavior of a DSPy module. In contrast, teleprompters are powerful optimizers (included in DSPy) that can learn to bootstrap and select effective prompts for the modules of any program.
The signature consists of a minimal description of the sub-task and one or more input questions that will be asked to the LM. It also explains the question?s answer that we expect from the LM. The teleprompters are automated prompting at a distance. They say that compared to others, DSPy requires very minimal labeling. It will bootstrap any intermediate labels required to support the user?s pipeline, which involves multiple complex steps.
As the DSPy framework differs significantly from other libraries, it is quite easy to justify when to use it based on our use case. Researchers say that this unified framework comes in handy to an NLP/ AI researcher or someone who is exploring new pipelines or new tasks to solve advanced and complex problems. To make it accessible to everyone, they have released an installation user manual. They also say that various introductory tutorials and demos with reference material will be released in the future.