legal-flan-t5-large-lora-2

Legal Flan-T5 Large (LoRA) - Indian Court Case Prediction

This model is a fine-tuned version of google/flan-t5-large using Low-Rank Adaptation (LoRA). It is specifically designed for the Indian legal domain to perform dual tasks: predicting the outcome of a legal case (Accepted/Rejected) and generating a concise summary or explanation of the judgment.

Model Details

Model Description

This model acts as the predictive engine for Legal Text Summariser (LTS) applications. It ingests excerpts from Indian court cases (front and tail sections of the document) and outputs a structured response containing a discrete label and a brief generative explanation.

• Developed by: Swastik Sharma (@Beelzi)
• Model type: Sequence-to-Sequence Language Model (PEFT/LoRA)
• Language(s) (NLP): English (Indian Legal Context)
• License: Apache 2.0 (Inherited from base model)
• Finetuned from model: google/flan-t5-large

Model Sources

• Repository: https://huggingface.co/Beelzi/legal-flan-t5-large-lora-2

Uses

Direct Use

The model is intended to be used by legal tech developers, researchers, and law students to:

1. Classify/Predict: Determine the likely outcome of a case (Label 0: Rejected, Label 1: Accepted).
2. Summarize: Provide a brief, legally coherent explanation for the predicted outcome based on the input text.

Downstream Use

It is optimized for integration into Retrieval-Augmented Generation (RAG) pipelines and decentralized security suites analyzing legal or compliance documents.

Out-of-Scope Use

• The model should not be used as a substitute for professional legal counsel.
• It is not designed for general-purpose conversational chat or creative writing.
• Performance on non-Indian legal jurisdictions (e.g., US or UK law) is untested and likely suboptimal.

Bias, Risks, and Limitations

• Domain Specificity: The model's reasoning is limited to the patterns present in the PredEx dataset.
• Context Window: Due to the 512-token limit of T5-Large, long legal documents must be truncated (e.g., extracting the introduction and conclusion) which may result in a loss of critical middle-text context.
• Automation Bias: Users should avoid over-relying on the model's predictions for real-world legal strategies without human verification.

How to Get Started with the Model

Use the code below to load the base model and the LoRA adapter for inference:

from transformers import T5ForConditionalGeneration, T5Tokenizer
from peft import PeftModel
import torch

# Load Base Model and Tokenizer
base_model_id = "google/flan-t5-large"
tokenizer = T5Tokenizer.from_pretrained(base_model_id)
base_model = T5ForConditionalGeneration.from_pretrained(base_model_id, device_map="auto")

# Load LoRA Adapter
peft_model_id = "Beelzi/legal-flan-t5-large-lora"
model = PeftModel.from_pretrained(base_model, peft_model_id)

# Inference Example
prompt = "Predict the legal outcome of this Indian court case. Output Label: 0 (rejected) or Label: 1 (accepted), then a brief explanation. Case: [INSERT CASE TEXT HERE]"
inputs = tokenizer(prompt, return_tensors="pt", max_length=512, truncation=True).to("cuda")

outputs = model.generate(**inputs, max_length=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))