metadata
license: mit
language:
- en
metrics:
- perplexity
- accuracy
base_model:
- microsoft/Phi-3-mini-4k-instruct
library_name: transformers
tags:
- legal
🧑⚖️ vakil-phi3-mini-4k-instruct-finetuned
📌 Overview
This repository hosts a fine‑tuned version of microsoft/Phi‑3‑mini‑4k‑instruct (3.8B parameters), adapted specifically for Indian legal knowledge tasks.
The model was instruction‑tuned using LoRA (Low‑Rank Adaptation) on curated datasets covering constitutional acts and statutory sections.
The objective of this fine‑tuning was to enhance the model’s ability to deliver accurate, contextual, and explainable outputs for legal queries in the Indian domain.
⚙️ Training Details
- Base Model: microsoft/Phi‑3‑mini‑4k‑instruct
- Fine‑Tuning Method: LoRA (parameter‑efficient fine‑tuning)
- Domain Data: Indian constitutional acts, statutory sections, and related legal texts
- Training Infrastructure: RunPod RTX A6000 GPU
- Training Duration: 18 hours, 2 epochs
- Optimization Goal: Reduce training loss and improve domain‑specific accuracy
📊 Evaluation
Intrinsic Evaluation:
- Reduced perplexity compared to the base model
- Improved accuracy on domain‑specific test sets
Extrinsic Evaluation:
- Better parsing of statutes and structured legal outputs
- Enhanced contextual reasoning in legal Q&A tasks
Qualitative Observations:
- More consistent responses when asked about constitutional provisions
- Improved ability to generate structured JSON outputs for legal sections
🚀 Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "ManjunathCode10x/vakil-phi3-mini-4k-instruct-finetuned"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)
inputs = tokenizer("Explain Article 21 of the Indian Constitution:", return_tensors="pt")
outputs = model.generate(**inputs, max_length=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))