Create chat.py
Browse files
chat.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import streamlit as st
|
3 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments
|
4 |
+
import torch
|
5 |
+
from torch.utils.data import Dataset
|
6 |
+
|
7 |
+
# Initialize model and tokenizer as global variables
|
8 |
+
model = None
|
9 |
+
tokenizer = None
|
10 |
+
|
11 |
+
# Dictionary to store user instructions for future responses
|
12 |
+
user_instructions = {}
|
13 |
+
|
14 |
+
# Dummy dataset class for user feedback
|
15 |
+
class FeedbackDataset(Dataset):
|
16 |
+
def __init__(self, input_texts, target_texts):
|
17 |
+
self.input_texts = input_texts
|
18 |
+
self.target_texts = target_texts
|
19 |
+
|
20 |
+
def __len__(self):
|
21 |
+
return len(self.input_texts)
|
22 |
+
|
23 |
+
def __getitem__(self, idx):
|
24 |
+
inputs = tokenizer.encode(self.input_texts[idx], return_tensors="pt").squeeze()
|
25 |
+
targets = tokenizer.encode(self.target_texts[idx], return_tensors="pt").squeeze()
|
26 |
+
return {"input_ids": inputs, "labels": targets}
|
27 |
+
|
28 |
+
def load_model(model_name_or_path):
|
29 |
+
global model, tokenizer
|
30 |
+
|
31 |
+
st.write(f"Loading model from {model_name_or_path}...")
|
32 |
+
|
33 |
+
# Load the tokenizer and model
|
34 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
|
35 |
+
model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
|
36 |
+
|
37 |
+
st.success("Model loaded successfully!")
|
38 |
+
|
39 |
+
def generate_response(input_text):
|
40 |
+
# Ensure model and tokenizer are loaded
|
41 |
+
if model is None or tokenizer is None:
|
42 |
+
st.error("Model is not loaded. Please load a model first.")
|
43 |
+
return ""
|
44 |
+
|
45 |
+
# Check if there's a user-defined response
|
46 |
+
if input_text in user_instructions:
|
47 |
+
return user_instructions[input_text]
|
48 |
+
|
49 |
+
# Encode input text
|
50 |
+
inputs = tokenizer.encode(input_text, return_tensors="pt")
|
51 |
+
|
52 |
+
# Generate response using the model
|
53 |
+
with torch.no_grad():
|
54 |
+
outputs = model.generate(
|
55 |
+
inputs, max_length=100, num_return_sequences=1, do_sample=True, top_k=50, top_p=0.95
|
56 |
+
)
|
57 |
+
|
58 |
+
# Decode and return the response
|
59 |
+
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
60 |
+
return response
|
61 |
+
|
62 |
+
def train_on_feedback(input_text, correct_response):
|
63 |
+
# Prepare dataset
|
64 |
+
dataset = FeedbackDataset([input_text], [correct_response])
|
65 |
+
|
66 |
+
# Training arguments
|
67 |
+
training_args = TrainingArguments(
|
68 |
+
output_dir="./feedback_model",
|
69 |
+
num_train_epochs=1,
|
70 |
+
per_device_train_batch_size=1,
|
71 |
+
learning_rate=1e-5,
|
72 |
+
logging_dir='./logs',
|
73 |
+
logging_steps=10,
|
74 |
+
save_steps=100
|
75 |
+
)
|
76 |
+
|
77 |
+
# Trainer for the feedback loop
|
78 |
+
trainer = Trainer(
|
79 |
+
model=model,
|
80 |
+
args=training_args,
|
81 |
+
train_dataset=dataset,
|
82 |
+
)
|
83 |
+
|
84 |
+
# Train model on the feedback
|
85 |
+
trainer.train()
|
86 |
+
|
87 |
+
def chat_interface():
|
88 |
+
st.title("🤖 Chat with AI")
|
89 |
+
|
90 |
+
# Input for model name or path
|
91 |
+
model_name_or_path = st.text_input("Enter model name or local path:", "gpt2")
|
92 |
+
|
93 |
+
# Button to load the model
|
94 |
+
if st.button("Load Model"):
|
95 |
+
load_model(model_name_or_path)
|
96 |
+
|
97 |
+
st.write("---")
|
98 |
+
|
99 |
+
# Chat input
|
100 |
+
input_text = st.text_input("You:")
|
101 |
+
|
102 |
+
if st.button("Send"):
|
103 |
+
response = generate_response(input_text)
|
104 |
+
st.write("AI:", response)
|
105 |
+
|
106 |
+
# Feedback section
|
107 |
+
feedback = st.radio("Was this response helpful?", ("Yes", "No"))
|
108 |
+
|
109 |
+
if feedback == "No":
|
110 |
+
correct_response = st.text_input("What should the AI have said?")
|
111 |
+
if st.button("Submit Feedback"):
|
112 |
+
# Train model on feedback
|
113 |
+
train_on_feedback(input_text, correct_response)
|
114 |
+
st.success("Feedback recorded. AI will improve based on this feedback.")
|
115 |
+
|
116 |
+
# Run chat interface
|
117 |
+
if __name__ == "__main__":
|
118 |
+
chat_interface()
|