phase-1-sft-legal-alligned

Model Description

LegalSafe-Falcon-7B is a safety-aligned large language model fine-tuned for the Indian legal domain. The model focuses on reducing toxic, biased, and unsafe outputs while maintaining high accuracy in legal reasoning and responses. It integrates Constitutional AI principles and Reinforcement Learning from AI Feedback (RLAIF) to improve ethical alignment.

• Developed by: Gopi M
• Model type: Causal Language Model (LLM)
• Language(s): English (Legal Domain - Indian Context)
• License: Apache 2.0 (inherits from base model)
• Finetuned from model: ybelkada/falcon-7b-sharded-bf16

---

Uses

Direct Use

• Legal question answering
• Legal document summarization
• Safe text generation in sensitive domains
• Educational tools for legal studies

Downstream Use

• Integration into legal chatbots
• AI-based legal assistants
• Compliance and advisory systems
• Safety-critical NLP applications

Out-of-Scope Use

• Providing official legal advice
• Use in high-risk legal decision-making without human oversight
• Generating harmful, biased, or illegal content

---

Bias, Risks, and Limitations

• May still exhibit residual bias from training data
• Performance may vary outside Indian legal domain
• Not a substitute for professional legal consultation
• Risk of hallucinations in complex legal scenarios

Recommendations

• Use with human oversight in legal applications
• Validate outputs before deployment in critical systems
• Avoid relying on model for final legal decisions

---

How to Get Started

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "gopi30/phase-1-sft-legal-alligned"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "Explain the concept of fundamental rights in Indian law."
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))