Constructing comprehensive court orders

[Gautam-Rajeev]

Goal

An agent that allows for a guided drafting of documents in government offices. Examples of documents include - court judgements, letters, office orders. It gathers data by querying the users in a conversational format and drafting a document in the style specific to the document.

Description

The court clerical staff currently take several days to draft a comprehensive court order even after they have collected all relevant information from all the required parties. Most judgements follow a predefined structure including summarizing/structuring all content (this doesn't require any higher reasoning usually and is very repetitive work). The chatbot aims to automate this effort.
Any required content could be sourced through a form or through a conversation that fills in any missing information.
The bot then would generate a judgement using the content in a similar style as the previous judgements and without stating any falsities or creating its own conclusions. This should ideally be completed in less than 10 minutes.

Implementation details

This will involve:

• Defining a generic structure of documents and creating a registry of document (documentType, documentForm, style, etc)
• Building a conversation agent that allows for gathering information needed to draft the document.
• Drafting the document
  • Fine-tuning an LLM to match the style of previous judgements
  • Dataset Generation based on older documents
  • Benchmarking LLMs on complex tasks like judgement validity

Students to be selected

1 students

Requirements

Python, Pytorch, LLM , NLP

[sarthak13agr]

Can we make this more expansive so that it covers almost all 'drafting' tasks that occur in a government offices? Besides court judgements, this will then include other things such as letters and office orders, The nature of the bot should be the same - a query tool to ask the user to fill in important details, after which it generates a document that is consistent in style, structure, and language with the human-written documents we have been using in the past.

Also, this goal - "Create a bot that is capable of asking questions that all relevant content required to construct a judgement and combine them meaningfully in the right structure into a comprehensive and valid court judgement" can be written better

[sarthak13agr]

"Fine-tuning an LLM to match the style of previous judgements
Providing multi-shot examples to match style
Being able to fine tune an LLM to carry out task
Benchmarking LLMs on complex tasks like judgement validity"

Are these repetitive?

[ChakshuGautam]

"Fine-tuning an LLM to match the style of previous judgements Providing multi-shot examples to match style Being able to fine tune an LLM to carry out task Benchmarking LLMs on complex tasks like judgement validity"

Are these repetitive?

• Fine-tuning an LLM to match the style of previous judgements - Only solving for style of text.
• Providing multi-shot examples to match style - Rewording this to - Dataset Generation based on older documents
• Being able to fine tune an LLM to carry out task - Removing
• Benchmarking LLMs on complex tasks like judgement validity - Keeping as is

[ChakshuGautam]

Can we make this more expansive so that it covers almost all 'drafting' tasks that occur in a government offices? Besides court judgements, this will then include other things such as letters and office orders, The nature of the bot should be the same - a query tool to ask the user to fill in important details, after which it generates a document that is consistent in style, structure, and language with the human-written documents we have been using in the past.

Also, this goal - "Create a bot that is capable of asking questions that all relevant content required to construct a judgement and combine them meaningfully in the right structure into a comprehensive and valid court judgement" can be written better

Incorporated.

[sarthak13agr]

Can we make this more expansive so that it covers almost all 'drafting' tasks that occur in a government offices? Besides court judgements, this will then include other things such as letters and office orders, The nature of the bot should be the same - a query tool to ask the user to fill in important details, after which it generates a document that is consistent in style, structure, and language with the human-written documents we have been using in the past.
Also, this goal - "Create a bot that is capable of asking questions that all relevant content required to construct a judgement and combine them meaningfully in the right structure into a comprehensive and valid court judgement" can be written better

Incorporated.

Goal looks good but can we edit the other sections too?

[ChakshuGautam]