metadata
license: mit
datasets:
- fka/awesome-chatgpt-prompts
language:
- en
metrics:
- accuracy
base_model:
- black-forest-labs/FLUX.1-Krea-dev
new_version: openai/gpt-oss-120b
pipeline_tag: token-classification
library_name: fastai
tags:
- code
Got it — you want the RL chatbot in a single Python file that:
Trains with simulated user feedback (or can take live user feedback)
Uses Reinforcement Learning (Q-learning) to improve
Is complete and runnable without external setup except pip install scikit-learn
Here’s a single-file complete working example:
import random
import pickle
import os
import numpy as np
from collections import defaultdict
from sklearn.feature_extraction.text import CountVectorizer
# ------------------------------
# RL Chatbot Agent (Q-Learning)
# ------------------------------
class RLChatbot:
def __init__(self, actions, alpha=0.1, gamma=0.9, epsilon=0.2):
self.actions = actions # possible responses
self.alpha = alpha # learning rate
self.gamma = gamma # discount factor
self.epsilon = epsilon # exploration rate
self.q_table = defaultdict(lambda: np.zeros(len(actions)))
self.vectorizer = CountVectorizer()
def featurize(self, text):
"""Convert input text to a hashed state ID (string key)."""
return str(hash(text.lower()) % (10**8))
def choose_action(self, state):
"""Epsilon-greedy action selection."""
if random.random() < self.epsilon:
return random.randint(0, len(self.actions) - 1)
return int(np.argmax(self.q_table[state]))
def update(self, state, action, reward, next_state):
"""Q-learning update."""
old_q = self.q_table[state][action]
next_max = np.max(self.q_table[next_state])
self.q_table[state][action] += self.alpha * (reward + self.gamma * next_max - old_q)
def save(self, path="rl_chatbot.pkl"):
with open(path, "wb") as f:
pickle.dump((dict(self.q_table), self.actions), f)
def load(self, path="rl_chatbot.pkl"):
if os.path.exists(path):
with open(path, "rb") as f:
data = pickle.load(f)
self.q_table = defaultdict(lambda: np.zeros(len(self.actions)), data[0])
self.actions = data[1]
# ------------------------------
# Simulated training environment
# ------------------------------
def simulated_reward(user_input, bot_response):
"""Fake reward function for simulation:
Higher reward if bot_response 'matches' intent."""
if "hello" in user_input.lower() and "hello" in bot_response.lower():
return 5
if "bye" in user_input.lower() and "bye" in bot_response.lower():
return 5
if "help" in user_input.lower() and "help" in bot_response.lower():
return 5
return -1 # default negative reward
# ------------------------------
# Main program
# ------------------------------
if __name__ == "__main__":
actions = [
"Hello! How can I help you?",
"Goodbye! Have a nice day.",
"I can help with your problems. What do you need?",
"I'm not sure I understand.",
"Please tell me more."
]
agent = RLChatbot(actions)
# ------------------------------
# Simulated training phase
# ------------------------------
training_data = [
"hello", "hi there", "bye", "goodbye", "i need help", "can you help me",
"what's up", "please help", "bye bye", "see you"
]
for episode in range(200):
user_msg = random.choice(training_data)
state = agent.featurize(user_msg)
action = agent.choose_action(state)
bot_reply = actions[action]
reward = simulated_reward(user_msg, bot_reply)
next_state = agent.featurize("end") # stateless
agent.update(state, action, reward, next_state)
print("✅ Training completed (simulated)")
# Save trained model
agent.save()
# ------------------------------
# Interactive chat
# ------------------------------
print("\n🤖 RL Chatbot is ready! Type 'quit' to exit.")
agent.load()
while True:
user_input = input("You: ")
if user_input.lower() in ["quit", "exit"]:
break
state = agent.featurize(user_input)
action = agent.choose_action(state)
bot_reply = actions[action]
print(f"Bot: {bot_reply}")
# Get human feedback (reward)
try:
reward = int(input("Rate this reply (-5 to 5): "))
except ValueError:
reward = 0 # default if invalid
next_state = agent.featurize("end")
agent.update(state, action, reward, next_state)
agent.save()
print("💾 Chatbot model updated and saved.")
How it works:
Actions: Predefined bot replies.
Q-learning: Learns which reply works best for each type of user input.
Training: First, simulated with fake reward rules (matching intent).
Interactive mode: You chat with it and give feedback (-5 to 5), it updates its policy instantly.
Run it:
pip install scikit-learn
python rl_chatbot.py
If you want, I can now make a neural network-based version (Deep Q-Learning) so the chatbot learns richer, free-form responses instead of a fixed set. Do you want me to upgrade it to DQN? That way it feels more like ChatGPT learning live from you.