app.py

import os
import time
import gradio as gr
from unsloth import FastLanguageModel
from transformers import TrainingArguments, TextStreamer
from datasets import load_dataset

# --- Environment Fixes ---
os.environ["OMP_NUM_THREADS"] = "1"

# --- Load model ---
model_name = "EpistemeAI/Episteme-gptoss-20b-RL"
print(f"Loading model: {model_name}")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    load_in_4bit=False,
    dtype=None,
    device_map="auto",
)

# --- Simple memory for ongoing chat ---
conversation_history = []
training_data = []
TRAIN_EVERY_N_PROMPTS = 8
training_status = "Idle"

# --- Function: Generate chat response ---
def chat_response(user_input):
    global training_data, training_status

    # Record the prompt
    conversation_history.append({"role": "user", "content": user_input})

    # Build prompt
    full_prompt = "\n".join([f"{msg['role']}: {msg['content']}" for msg in conversation_history])
    inputs = tokenizer(full_prompt, return_tensors="pt").to(model.device)

    # Stream response
    streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    model.generate(**inputs, streamer=streamer, max_new_tokens=200)
    response_text = streamer.text.strip()

    # Store assistant message
    conversation_history.append({"role": "assistant", "content": response_text})
    training_data.append({"instruction": user_input, "expected_output": response_text})

    # Auto-train trigger
    if len(training_data) >= TRAIN_EVERY_N_PROMPTS:
        training_status = "Training..."
        yield [
            conversation_history,
            "Auto-training in progress... this may take a moment."
        ]
        train_model(training_data)
        training_data.clear()
        training_status = "Idle"

    yield [conversation_history, f"Assistant: {response_text}"]

# --- Function: Train model on recent conversations ---
def train_model(data_samples):
    dataset = {
        "instruction": [d["instruction"] for d in data_samples],
        "expected_output": [d["expected_output"] for d in data_samples],
    }
    training_args = TrainingArguments(
        output_dir="./results",
        per_device_train_batch_size=2,
        num_train_epochs=1,
        learning_rate=2e-4,
        logging_dir="./logs",
        save_strategy="no",
        optim="adamw_bnb_8bit",
        fp16=True,
        lr_scheduler_type="cosine",
    )

    model.train_model(
        dataset=dataset,
        training_args=training_args,
        lora_r=8,
        lora_alpha=16,
        lora_dropout=0.1,
    )
    print("✅ Model retrained successfully")

# --- Function: Display training status ---
def get_training_status():
    return f"Training status: {training_status}"

# --- Function: Clear chat ---
def clear_chat():
    global conversation_history
    conversation_history = []
    return [], "Chat cleared."

# --- Build Gradio UI ---
with gr.Blocks(title="🧠 Auto-Training Chatbot", theme=gr.themes.Soft()) as demo:
    gr.Markdown("### 🤖 Adaptive AI Chatbot (with Auto-Retraining)")

    with gr.Row():
        chatbot = gr.Chatbot(
            height=500,
            label="Chat Interface",
            type="messages"
        )
        with gr.Column():
            status_display = gr.Textbox(
                label="Training Status",
                value="Idle",
                interactive=False
            )
            clear_btn = gr.Button("Clear Chat")

    user_input = gr.Textbox(
        label="Your message",
        placeholder="Type your question or message here..."
    )
    send_btn = gr.Button("Send")

    # Bind events
    send_btn.click(
        fn=chat_response,
        inputs=[user_input],
        outputs=[chatbot, status_display],
    )

    clear_btn.click(
        fn=clear_chat,
        outputs=[chatbot, status_display],
    )

    # Add auto-refresh every 2s
    timer = gr.Timer(interval=2.0, active=True)
    timer.tick(fn=get_training_status, outputs=[status_display])

# --- Launch ---
if __name__ == "__main__":
    demo.launch()