app.py

import os
import gradio as gr
from transformers import pipeline, AutoTokenizer
import torch
import spaces
import json
from huggingface_hub import HfApi, upload_file

# --- Constants ---
HF_TOKEN = os.environ.get("HF_TOKEN")
DATASET_REPO = "Pisethan/khmer-lesson-dataset-generated"
LOCAL_JSONL = "generated_lessons.jsonl"

# --- Options ---
grade_options = ["1", "2", "3", "4", "5", "6"]
topic_options = ["Addition", "Subtraction", "Counting", "Number Recognition", "Multiplication", "Division"]
level_options = ["Beginner", "Intermediate", "Advanced"]

# --- Tokenizer (global) ---
tokenizer = AutoTokenizer.from_pretrained("Pisethan/khmer-lesson-model", token=HF_TOKEN)

# --- Helper to save and upload ---
def save_to_jsonl(record):
    with open(LOCAL_JSONL, "a", encoding="utf-8") as f:
        f.write(json.dumps(record, ensure_ascii=False) + "\n")

    upload_file(
        path_or_fileobj=LOCAL_JSONL,
        path_in_repo="generated_lessons.jsonl",
        repo_id=DATASET_REPO,
        repo_type="dataset",
        token=HF_TOKEN
    )

# --- Generation for one lesson ---
@spaces.GPU
def generate_lesson(grade, topic, level):
    device = 0 if torch.cuda.is_available() else -1
    pipe = pipeline(
        "text-generation",
        model="Pisethan/khmer-lesson-model-v2",
        tokenizer=tokenizer,
        device=device,
        token=HF_TOKEN
    )

    prompt = f"""
You are a lesson planning assistant. Return only one structured Khmer math lesson plan with these fields:

Lesson Title:
Objective:
Activity:
Instruction (Khmer):
Materials:

Please follow the structure exactly.

Grade: {grade}
Topic: {topic}
TaRL Level: {level}
"""

    output = pipe(prompt, max_new_tokens=300, temperature=0.7, do_sample=True, eos_token_id=tokenizer.eos_token_id)
    result = output[0]['generated_text']

    # Save to dataset
    record = {
        "grade": grade,
        "topic": topic,
        "level": level,
        "prompt": prompt.strip(),
        "completion": result.strip()
    }
    save_to_jsonl(record)
    return result

# --- Generation for all combinations ---
@spaces.GPU
def generate_all_lessons():
    device = 0 if torch.cuda.is_available() else -1
    pipe = pipeline(
        "text-generation",
        model="Pisethan/khmer-lesson-model-v2",
        tokenizer=tokenizer,
        device=device,
        token=HF_TOKEN
    )

    results = ""
    for grade in grade_options:
        for topic in topic_options:
            for level in level_options:
                prompt = f"""Generate a Khmer math lesson plan.

Grade: {grade}
Topic: {topic}
TaRL Level: {level}"""
                output = pipe(prompt, max_new_tokens=200, temperature=0.7, do_sample=True)
                result = output[0]['generated_text']

                record = {
                    "grade": grade,
                    "topic": topic,
                    "level": level,
                    "prompt": prompt.strip(),
                    "completion": result.strip()
                }
                save_to_jsonl(record)

                results += f"🔹 ថ្នាក់ {grade} | {topic} | {level}\n{result}\n\n{'-'*50}\n\n"
    return results

# --- UI ---
with gr.Blocks() as demo:
    gr.Markdown("## 🤖 អ្នកជំនួយបង្កើតមេរៀនគណិតវិទ្យា")
    gr.Markdown("ជ្រើសរើសថ្នាក់ ប្រធានបទ និងកម្រិតសិស្ស រួចចុចបង្កើតមេរៀន។ ឬចុចប៊ូតុងខាងក្រោមសម្រាប់បង្កើតមេរៀនទាំងអស់។")

    with gr.Row():
        grade = gr.Dropdown(choices=grade_options, label="ថ្នាក់ (Grade)", value="1")
        topic = gr.Dropdown(choices=topic_options, label="ប្រធានបទ (Topic)", value="Addition")
        level = gr.Dropdown(choices=level_options, label="កម្រិតសិស្ស (TaRL Level)", value="Beginner")

    output_box = gr.Textbox(
        label="📘 Khmer Lesson Plan",
        lines=20,
        max_lines=200,
        show_copy_button=True,
        autoscroll=True
    )

    with gr.Row():
        gen_btn = gr.Button("✅ បង្កើតមេរៀន")
        gen_all_btn = gr.Button("🧠 បង្កើតមេរៀនទាំងអស់")
        clear_btn = gr.Button("🧹 សម្អាត")

    gen_btn.click(fn=generate_lesson, inputs=[grade, topic, level], outputs=output_box)
    gen_all_btn.click(fn=generate_all_lessons, outputs=output_box)
    clear_btn.click(fn=lambda: "", outputs=output_box)

demo.queue()
demo.launch()