Upgrade VakilAI chat interface
Browse files
app.py
CHANGED
|
@@ -4,36 +4,76 @@ import torch
|
|
| 4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 5 |
from peft import PeftModel
|
| 6 |
|
|
|
|
| 7 |
BASE_MODEL = "unsloth/llama-3.2-3b-bnb-4bit"
|
|
|
|
|
|
|
| 8 |
ADAPTER_MODEL = "devNaam/vakilai-llama32-3b-v1"
|
| 9 |
|
|
|
|
| 10 |
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
|
| 11 |
|
|
|
|
| 12 |
model = AutoModelForCausalLM.from_pretrained(
|
| 13 |
BASE_MODEL,
|
| 14 |
device_map="auto"
|
| 15 |
)
|
| 16 |
|
|
|
|
| 17 |
model = PeftModel.from_pretrained(model, ADAPTER_MODEL)
|
| 18 |
|
| 19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 22 |
|
| 23 |
output = model.generate(
|
| 24 |
**inputs,
|
| 25 |
-
max_new_tokens=
|
| 26 |
-
temperature=0.
|
|
|
|
| 27 |
)
|
| 28 |
|
| 29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
|
| 31 |
|
| 32 |
-
|
|
|
|
| 33 |
fn=vakil_ai,
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
)
|
| 38 |
|
| 39 |
demo.launch()
|
|
|
|
| 4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 5 |
from peft import PeftModel
|
| 6 |
|
| 7 |
+
# Base model used during training
|
| 8 |
BASE_MODEL = "unsloth/llama-3.2-3b-bnb-4bit"
|
| 9 |
+
|
| 10 |
+
# Your VakilAI LoRA adapter
|
| 11 |
ADAPTER_MODEL = "devNaam/vakilai-llama32-3b-v1"
|
| 12 |
|
| 13 |
+
print("Loading tokenizer...")
|
| 14 |
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
|
| 15 |
|
| 16 |
+
print("Loading base model...")
|
| 17 |
model = AutoModelForCausalLM.from_pretrained(
|
| 18 |
BASE_MODEL,
|
| 19 |
device_map="auto"
|
| 20 |
)
|
| 21 |
|
| 22 |
+
print("Loading VakilAI adapter...")
|
| 23 |
model = PeftModel.from_pretrained(model, ADAPTER_MODEL)
|
| 24 |
|
| 25 |
+
print("Model ready.")
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
# Prompt template for legal assistant behavior
|
| 29 |
+
def build_prompt(user_question):
|
| 30 |
+
return f"""
|
| 31 |
+
You are VakilAI, an AI legal assistant specializing in Indian law.
|
| 32 |
+
|
| 33 |
+
Explain legal concepts clearly in simple language.
|
| 34 |
+
If possible, mention relevant IPC sections or legal principles.
|
| 35 |
+
|
| 36 |
+
User Question:
|
| 37 |
+
{user_question}
|
| 38 |
+
|
| 39 |
+
Answer:
|
| 40 |
+
"""
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
# Generate response
|
| 44 |
+
def vakil_ai(user_message, history):
|
| 45 |
+
|
| 46 |
+
prompt = build_prompt(user_message)
|
| 47 |
|
| 48 |
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 49 |
|
| 50 |
output = model.generate(
|
| 51 |
**inputs,
|
| 52 |
+
max_new_tokens=250,
|
| 53 |
+
temperature=0.5,
|
| 54 |
+
top_p=0.9
|
| 55 |
)
|
| 56 |
|
| 57 |
+
response = tokenizer.decode(output[0], skip_special_tokens=True)
|
| 58 |
+
|
| 59 |
+
# Extract only the answer part
|
| 60 |
+
if "Answer:" in response:
|
| 61 |
+
response = response.split("Answer:")[-1].strip()
|
| 62 |
+
|
| 63 |
+
return response
|
| 64 |
|
| 65 |
|
| 66 |
+
# Chat interface
|
| 67 |
+
demo = gr.ChatInterface(
|
| 68 |
fn=vakil_ai,
|
| 69 |
+
title="⚖️ AI Vakil – Indian Legal Assistant",
|
| 70 |
+
description="Ask questions about Indian law, IPC sections, and legal concepts.",
|
| 71 |
+
examples=[
|
| 72 |
+
"What is IPC Section 307?",
|
| 73 |
+
"What is the punishment for theft in India?",
|
| 74 |
+
"What is the difference between murder and culpable homicide?",
|
| 75 |
+
"What rights does a person have during arrest in India?"
|
| 76 |
+
]
|
| 77 |
)
|
| 78 |
|
| 79 |
demo.launch()
|