main.py

from fastapi import FastAPI, Request
from fastapi.staticfiles import StaticFiles
from fastapi.responses import FileResponse
from fastapi import FastAPI, BackgroundTasks, HTTPException, Query
from fastapi.responses import StreamingResponse
from starlette.concurrency import run_in_threadpool
from datasets import load_dataset
import random
import json
from genson import SchemaBuilder
from pathvalidate import sanitize_filename
from openai import OpenAI
import hashlib
import base64
from pprint import pprint
import asyncio
import importlib.util
import traceback
import sys
import json
import jsonschema
from utils import extract_code
import numpy as np
import os
import requests
import secrets
import urllib.parse

app = FastAPI()

client_id = os.getenv("OAUTH_CLIENT_ID")
client_secret = os.getenv("OAUTH_CLIENT_SECRET")
space_host = os.getenv("SPACE_HOST")

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ.get('OPENROUTER_KEY')
)

state_queue_map = {}

def is_sharegpt(sample):
    schema = {
        '$schema': 'http://json-schema.org/schema#',
        'type': 'object',
        'properties': {
            'conversations': {
                'type': 'array',
                'items': {
                    'type': 'object',
                    'properties': {
                        'from': {
                            'type': 'string',
                            'enum': [
                                'human',
                                'gpt',
                                'system']},
                        'value': {
                            'type': 'string'}},
                    'required': [
                        'from',
                        'value']}}},
        'required': ['conversations']}
    try:
        jsonschema.validate(instance=sample, schema=schema)
        return True
    except jsonschema.exceptions.ValidationError as e:
        return False


def sha256(string):
    # Create a hashlib object for SHA-256
    sha256_hash = hashlib.sha256()
    # Update the hash object with your string encoded as bytes
    sha256_hash.update(string.encode('utf-8'))

    return sha256_hash.hexdigest()


def get_adapter_name(sample):
    builder = SchemaBuilder()
    builder.add_object(sample)
    schema = builder.to_schema()

    return sha256(json.dumps(schema))


def has_adapter(sample):
    adapter_name = get_adapter_name(sample)

    module_name = f"dataset_adapters.{adapter_name}"
    module_spec = importlib.util.find_spec(module_name)

    if module_spec is None:
        return False

    return True


def auto_tranform(sample):
    adapter_name = get_adapter_name(sample)
    if not has_adapter(sample):
        create_adapter(sample, adapter_name)

    module_name = f"dataset_adapters.{adapter_name}"
    spec = importlib.util.spec_from_file_location(
        module_name, f"dataset_adapters/{adapter_name}.py")
    dynamic_module = importlib.util.module_from_spec(spec)
    sys.modules[module_name] = dynamic_module
    spec.loader.exec_module(dynamic_module)

    transformed_data = dynamic_module.transform_data(sample)

    if isinstance(transformed_data, list):
        return {'conversations': transformed_data}

    return transformed_data

    with open(f"dataset_adapters/{adapter_name}.py", 'w') as file:
        file.write(code_string)


def create_adapter(sample, adapter_name):
    builder = SchemaBuilder()
    builder.add_object(sample)
    schema = builder.to_schema()

    prompt = f"""Make me minimal and efficient python code to convert data in the shape of

initial data shape
==========➡️📑📐==========
```jsonschema
{schema}
```
==========➡️📑📐==========

to equivalent data in the form

final data shape
==========⬇️📑📐==========
```jsonschema
{{'$schema': 'http://json-schema.org/schema#', 'type': 'object', 'properties': {{'conversations': {{'type': 'array', 'items': {{'type': 'object', 'properties': {{'from': {{ 'type': 'string', 'enum': ['human', 'gpt', 'system'] }}, 'value': {{'type': 'string'}}}}, 'required': ['from', 'value']}}}}}}, 'required': ['conversations']}}
```
==========⬇️📑📐==========

the data to transform is
```json
{sample}
```


Inside the data to transform, `input` and `instruction` is usually associated with `"from" : "human"` while `output` is usually associated with `"from" : "gpt"`

For transforming the data you shall use python. Make robust and elegant python code that will do the transformation


your code will contain a function `def transform_data(data):` that does the transformation and outputs the newly shaped data. Only the data, no schema. Your code snippet will include only the function signature and body. I know how to call it. You won't need to import anything, I will take care of parsing and dumping json. You work with dicts. Remember to be careful if you iterate over the data because I want the output conversation to always start with the prompt. In other words, always process "input" before "output" and "instruction" before "output". Such heuristics are very important. If there is "instruction" and "input" and the "input" is not empty, concat the input at the end of the first message. If the data contains no "system" message, human always speaks first. If it contains a "system"  message, the "system" message is first, then human, then gpt, then alternating if needed

"human" ALWAYS SPEAKS BEFORE "gpt", if you suspect your code makes "gpt speak first, fix it

MOST IMPORTANT IS THAT YOU look at the initial data shape (➡️📑📐) to ground your transformation into final data shape (⬇️📑📐)

Your output should contain only the code for `def transform_data(data):`, signature and body. Put the code inside markdown code block"""

    response = client.chat.completions.create(
        model="openai/gpt-4-1106-preview",
        messages=[
            {"role": "system", "content": """You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture.
Knowledge cutoff: 2023-04
Current date: 2023-11-05

Image input capabilities: Enabled"""},
            {"role": "user", "content": prompt}
        ]
    )

    val = response.choices[0].message.content

    code_string = extract_code(val)

    if code_string is None:
        raise Exception("hey la")

    with open(f"dataset_adapters/{adapter_name}.py", 'w') as file:
        file.write(code_string)


@app.get("/sample")
async def get_sample(hash: str = Query(..., alias="hash")):
    res = await get_sample_by_hash(hash)
    if res is None:
        raise HTTPException(status_code=404, detail="Item not found")
    data, dataset = res
    sample = auto_tranform(json.loads(data))
    return {'sample': sample, 'dataset': dataset}


def generate_random_string(length=16):
    return secrets.token_hex(length)



@app.get("/oauth_token")
async def get_oauth_token():
    queue = asyncio.Queue()

    async def event_stream(queue, state):
        state_queue_map[state] = queue

        redirect_uri = f'https://{space_host}/login/callback'

        auth_url = (
            f"https://huggingface.co/oauth/authorize?"
            f"redirect_uri={urllib.parse.quote(redirect_uri)}&"
            f"client_id={client_id}&"
            f"scope=openid%20profile&"
            f"response_type=code&"
            f"state={state}"
        )
        yield f"data: {json.dumps({ 'url' : auth_url })}\n\n"

        try:
            while True:
                message = await queue.get()
                if 'end_stream' in message and message['end_stream']:
                    break
                yield f"data: {json.dumps(message)}\n\n"
        finally:
            del state_queue_map[state]

    state = generate_random_string()
    return StreamingResponse(
        event_stream(queue, state),
        media_type="text/event-stream")


@app.get("/random-sample-stream")
async def get_random_sample(background_tasks: BackgroundTasks, dataset_name: str = Query(..., alias="dataset-name"), index: str = Query(None, alias="index")):
    queue = asyncio.Queue()

    def event_stream(queue):
        yield f"data: {json.dumps({'status': 'grab_sample'})}\n\n"
        try:
            headers = {"Authorization": f"Bearer {os.environ.get('HF_TOKEN')}"}
            API_URL = f"https://datasets-server.huggingface.co/info?dataset={dataset_name}"

            def query():
                response = requests.get(API_URL, headers=headers)
                return response.json()
            data = query()

            splits = data['dataset_info']['default']['splits']
            split = next(iter(splits.values()))

            num_samples = split['num_examples']
            split_name = split['name']

            idx = random.randint(
                0, num_samples) if index is None else int(index)

            API_URL = f"https://datasets-server.huggingface.co/rows?dataset={dataset_name}&config=default&split=train&offset={idx}&length=1"

            def query():
                headers = {
                    "Authorization": f"Bearer {os.environ.get('HF_TOKEN')}"}
                response = requests.get(API_URL, headers=headers)

                if response.status_code != 200:
                    raise Exception("hugging face api error")
                return response.json()
            data = query()

            random_sample = data['rows'][0]['row']

            hashed = sha256(json.dumps(random_sample))

        except Exception as e:
            message = ""
            if hasattr(e, 'message'):
                message = e.message
            else:
                message = str(e)

            print("error : ", message)
            yield f"data: {json.dumps({'status': 'error', 'message' : message })}\n\n"

        transformed_data = random_sample

        success = True

        if not is_sharegpt(random_sample):
            try:
                if not has_adapter(random_sample):
                    yield f"data: {json.dumps({'status': 'creating_adapter'})}\n\n"

                transformed_data = auto_tranform(random_sample)
            except Exception as e:
                success = False
                if hasattr(e, 'message'):
                    print("error : ", e.message)
                else:
                    print("error : ", e)
                yield f"data: {json.dumps({'status': 'error'})}\n\n"

        if success:
            yield f"data: {json.dumps({'status': 'done', 'data' : transformed_data, 'index' : str(idx)})}\n\n"

    return StreamingResponse(
        event_stream(queue),
        media_type="text/event-stream")


@app.get("/random-sample")
async def get_random_sample(dataset_name: str = Query(..., alias="dataset-name")):
    try:
        dataset = load_dataset(dataset_name, streaming=True)
        split = [key for key in dataset.keys() if "train" in key]
        dataset = load_dataset(dataset_name, split=split[0], streaming=True)

        buffer_size = 100  # Define a reasonable buffer size
        samples_buffer = [
            sample for _, sample in zip(
                range(buffer_size), dataset)]

        random_sample = random.choice(samples_buffer)

        hashed = sha256(json.dumps(random_sample))

        sanitized = sanitize_filename(dataset_name)
        module_name = f"dataset_adapters.{sanitized}"
        module_spec = importlib.util.find_spec(module_name)

        if module_spec is None:
            create_adapter(random_sample, sanitized)

        spec = importlib.util.spec_from_file_location(
            module_name, f"dataset_adapters/{sanitized}.py")
        dynamic_module = importlib.util.module_from_spec(spec)
        sys.modules[module_name] = dynamic_module
        spec.loader.exec_module(dynamic_module)

        transformed_data = dynamic_module.transform_data(random_sample)

        return transformed_data

    except FileNotFoundError:
        raise HTTPException(status_code=404, detail="Dataset not found")
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))


@app.get("/login/callback")
async def oauth_callback(code: str, state: str):
    credentials = f"{client_id}:{client_secret}"
    credentials_bytes = credentials.encode("ascii")
    base64_credentials = base64.b64encode(credentials_bytes)
    auth_header = f"Basic {base64_credentials.decode('ascii')}"
    username = ""

    try:
        token_response = requests.post(
            'https://huggingface.co/oauth/token',
            headers={'Authorization': auth_header},
            data={
                'grant_type': 'authorization_code',
                'code': code,
                'redirect_uri': f'https://{space_host}/login/callback',
                'client_id': client_id
            }
        )
        print(token_response.status_code, token_response.text)

        if token_response.status_code == 200:
            tokens = token_response.json()
            access_token = tokens.get('access_token')

            if access_token:
                url = "https://huggingface.co/oauth/userinfo"

                payload = ""
                headers = {
                    "Content-Type": "application/json",
                    "Authorization": f"Bearer {access_token}"
                }

                response = requests.request(
                    "GET", url, data=payload, headers=headers)

                print(response.text)

                if state in state_queue_map:
                    queue = state_queue_map[state]
                    await queue.put({"access_token": access_token, "username": username})
                    await queue.put({"end_stream": True})
            else:
                username = ""
        else:
            access_token = ""

    except Exception:
        traceback.print_exc()
        access_token = ""

    return {"access_token": access_token, "username": username}


@app.get("/oauth-config")
async def get_oauth_config(request: Request):
    return {
        "client_id": client_id,
        "redirect_uri": f'https://{space_host}/login/callback'
    }


@app.get("/")
def index() -> FileResponse:
    return FileResponse(path="static/index.html", media_type="text/html")


app.mount("/", StaticFiles(directory="static"), name="static")