⚖️ Indian Legal Opposing Counsel

A fine-tuned Qwen2.5-7B-Instruct model specialized as an opposing counsel for Indian law students to practice court arguments and deepen their understanding of the Indian Constitution.

🎯 What This Model Does

This model acts as a rigorous opposing counsel that:

Challenges legal arguments using the Indian Constitution (448 Articles, 12 Schedules)
Applies Legal Syllogism: Major Premise (law) → Minor Premise (facts) → Conclusion (argument)
Cites landmark precedents: Kesavananda Bharati, Maneka Gandhi, Vishaka, and other Supreme Court judgments
Presents counter-arguments with alternative interpretations, reasonable restrictions, and procedural challenges
Educates while arguing: Highlights strengths and suggests areas for improvement

🏗️ Training Details

Research Foundation

This model's training recipe is based on published legal AI research:

Paper	Contribution
INLegalLlama	LoRA SFT hyperparameters for Indian legal domain (~90% F1)
SaulLM-7B	Legal domain adaptation with synthetic conversations
DISC-LawLLM	Legal syllogism reasoning pattern
AgentCourt	Adversarial courtroom simulation approach

Datasets (26,326 combined examples)

Dataset	Rows	Content
viber1/indian-law-dataset	24,607	Writs, PIL, civil procedure, constitutional law, IPC
nisaar/Lawyer_GPT_India	150	Landmark cases, IPC, contract law, constitutional principles
RMani1/indian-legal-dataset-indian-law	1,569	Indian statutes, acts, legal provisions

Hyperparameters

Parameter	Value
Base model	Qwen/Qwen2.5-7B-Instruct
Method	LoRA SFT (Parameter-Efficient Fine-Tuning)
LoRA rank	32
LoRA alpha	64
LoRA target	all-linear layers
Learning rate	1e-4 (cosine schedule)
Effective batch size	16 (2 × 8 accumulation)
Epochs	3
Max sequence length	2048
Precision	bfloat16
Loss	assistant-only (system/user prompts masked)
Optimizer	AdamW fused

💻 Quick Start

Installation

pip install transformers peft torch

Inference

from transformers import pipeline

pipe = pipeline("text-generation", model="pkheria7/indian-legal-opposing-counsel")

messages = [
    {"role": "system", "content": "You are an experienced opposing counsel specializing in the Indian Constitution."},
    {"role": "user", "content": """I argue that my client's right to privacy under Article 21 of the Indian Constitution 
    has been violated by the government's surveillance program. The Supreme Court in Justice K.S. Puttaswamy v. 
    Union of India (2017) established that privacy is a fundamental right. Therefore, this surveillance is unconstitutional."""}
]

response = pipe(messages, max_new_tokens=512, do_sample=True, temperature=0.7)
print(response[0]["generated_text"][-1]["content"])

With PEFT (LoRA adapter)

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained("pkheria7/indian-legal-opposing-counsel")
tokenizer = AutoTokenizer.from_pretrained("pkheria7/indian-legal-opposing-counsel")

messages = [
    {"role": "system", "content": "You are an experienced opposing counsel specializing in the Indian Constitution."},
    {"role": "user", "content": "I argue that Section 377 IPC violates Article 14 and Article 21..."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs.to(model.device), max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

🏋️ Train It Yourself

On Google Colab (A100 recommended)

# Install dependencies
pip install torch transformers trl peft datasets accelerate bitsandbytes flash-attn

# Clone and run
git clone https://huggingface.co/pkheria7/indian-legal-opposing-counsel
cd indian-legal-opposing-counsel
python train.py

For T4 GPUs (16GB VRAM) — Use QLoRA

Edit train.py and set:

USE_QLORA = True

Hardware Requirements

GPU	VRAM	Mode	Est. Training Time
A100 80GB	80 GB	Full LoRA + Flash Attn	~2-3 hours
A100 40GB	40 GB	Full LoRA	~3-4 hours
A10G 24GB	24 GB	QLoRA (4-bit)	~4-5 hours
T4 16GB	16 GB	QLoRA (4-bit)	~6-8 hours

📊 Example Interactions

Example 1: Privacy Rights

Student: "Article 21 guarantees right to privacy. The Puttaswamy judgment says surveillance is unconstitutional."

Opposing Counsel: "While I acknowledge the landmark nature of Puttaswamy, counsel's argument oversimplifies the Court's holding. The nine-judge bench in Puttaswamy explicitly held that the right to privacy is NOT absolute — it is subject to reasonable restrictions satisfying the triple test: (1) legality, (2) legitimate aim, and (3) proportionality. The surveillance program satisfies all three..."

Example 2: Reservation Policy

Student: "Article 15(4) and Article 16(4) permit reservations. The 50% ceiling from Indra Sawhney is arbitrary."

Opposing Counsel: "Counsel's characterization of the 50% ceiling as 'arbitrary' ignores the constitutional reasoning in Indra Sawhney v. Union of India (1992). The Supreme Court held that exceeding 50% would violate Article 14's guarantee of equality..."

⚠️ Limitations

Not a substitute for legal advice: This model is for educational purposes only
Knowledge cutoff: The model's legal knowledge comes from training data and may not reflect the latest amendments or judgments
May hallucinate citations: Always verify case citations and article references independently
Dataset size: Training data is ~26K examples — more data (especially from IndianKanoon.org court judgments) would improve quality
Single-turn focus: The current training data is primarily single-turn Q&A; multi-turn adversarial debates would improve courtroom simulation