app.py

import numpy as np
import json
import gradio as gr

# === LOAD CORPUS & DATASET ===
with open("corpus.txt", "r", encoding="utf-8") as f:
    corpus = f.read().splitlines()

with open("dataset.json", "r", encoding="utf-8") as f:
    qa_data = json.load(f)

# === BUILD VOCAB ===
all_texts = corpus + list(qa_data.keys()) + list(qa_data.values())
vocab = list(set(" ".join(all_texts).split()))
word2id = {w: i for i, w in enumerate(vocab)}
id2word = {i: w for w, i in word2id.items()}
vocab_size = len(vocab)

def one_hot(word):
    vec = np.zeros(vocab_size)
    if word in word2id:
        vec[word2id[word]] = 1
    return vec

# === PARAMETER MODEL ===
np.random.seed(42)
hidden_size = 512   # bisa diubah
W1 = np.random.randn(vocab_size, hidden_size) * 0.01
W2 = np.random.randn(hidden_size, vocab_size) * 0.01
lr = 0.05

def softmax(x):
    e = np.exp(x - np.max(x))
    return e / e.sum()

def train_step(sentence):
    global W1, W2
    words = sentence.split()
    loss = 0
    for i in range(len(words)-1):
        x = one_hot(words[i])
        y_true = one_hot(words[i+1])
        h = np.dot(x, W1)
        o = np.dot(h, W2)
        y_pred = softmax(o)
        loss += np.mean((y_true - y_pred)**2)
        grad_o = y_pred - y_true
        dW2 = np.outer(h, grad_o)
        dW1 = np.outer(x, np.dot(W2, grad_o))
        W1 -= lr * dW1
        W2 -= lr * dW2
    return loss

# === PRETRAIN DENGAN CORPUS ===
for epoch in range(200):
    total_loss = 0
    for line in corpus:
        total_loss += train_step(line + " <END>")
    if epoch % 50 == 0:
        print(f"Pretrain Epoch {epoch}, Loss: {total_loss:.4f}")

# === FINE-TUNE DENGAN Q&A ===
for epoch in range(200):
    total_loss = 0
    for q, a in qa_data.items():
        total_loss += train_step(q + " " + a)
    if epoch % 50 == 0:
        print(f"Finetune Epoch {epoch}, Loss: {total_loss:.4f}")

# === GENERATE JAWABAN ===
def generate_reply(question, max_len=30):
    words = question.split()
    if words[0] not in word2id:
        return "Maaf, aku belum ngerti kata itu 🥺"
    x = one_hot(words[0])
    reply = []
    for _ in range(max_len):
        h = np.dot(x, W1)
        o = np.dot(h, W2)
        y_pred = softmax(o)
        pred_id = np.argmax(y_pred)  # argmax biar konsisten
        pred_word = id2word[pred_id]
        if pred_word == "<END>":
            break
        reply.append(pred_word)
        x = one_hot(pred_word)
    return " ".join(reply)

# === GRADIO INTERFACE ===
def chatbot(input_text):
    return generate_reply(input_text)

demo = gr.Interface(fn=chatbot,
                    inputs="text",
                    outputs="text",
                    title="Chatbot Numpy ala Cici 🤭",
                    description="Mini chatbot dengan training 2 tahap: corpus + Q&A")

if __name__ == "__main__":
    demo.launch()