git add -A && git commit -m "Update project ($(date -u +'%Y-%m-%dT%H:%M:%SZ'))" && git push origin main

# Uncomment to install the required packages
# %pip install -U together datasets transformers python-dotenv -q

import os
import together
from dotenv import load_dotenv

load_dotenv()

# Load API keys from environment variables
os.environ["TOGETHER_API_KEY"] = os.environ.get("TOGETHER_API_KEY", "")
os.environ["HUGGINGFACE_ACCESS_TOKEN"] = os.environ.get("HUGGINGFACE_ACCESS_TOKEN", "")
TOGETHER_API_KEY = os.environ.get("TOGETHER_API_KEY")
HUGGINGFACE_ACCESS_TOKEN = os.environ.get("HUGGINGFACE_ACCESS_TOKEN")

import json
from datasets import load_dataset

# Load a sample of 500 examples from the dataset
dataset = load_dataset("databricks/databricks-dolly-15k", split="train", token=HUGGINGFACE_ACCESS_TOKEN).select(range(500))

def format_for_finetuning(example):
    # Use a standard instruction format
    return {"text": f"<s>[INST] {example['instruction']} [/INST] {example['response']} </s>"}

formatted_dataset = dataset.map(format_for_finetuning)

# Save the prepared data to a JSONL file (only the 'text' field per line)
file_name = "dolly_prepared.jsonl"
with open(file_name, 'w') as f:
    for item in formatted_dataset:
        f.write(json.dumps({"text": item["text"]}) + "\n")

print(f"Dataset prepared and saved to {file_name}")
print("--- Sample Entry ---")
with open(file_name, 'r') as f:
    print(json.loads(f.readline())['text'])

# 1. Upload the training file
try:
    upload_response = together.Files.upload(file=file_name)
    training_file_id = upload_response['id']
    print(f"File uploaded successfully. File ID: {training_file_id}")

    # 2. Create the fine-tuning job
    fine_tune_response = together.Finetune.create(
      training_file=training_file_id,
      model='togethercomputer/llama-2-7b-chat', # Base model to fine-tune
      n_epochs=3,                            # Number of training epochs
      n_checkpoints=1,                       # Number of checkpoints to save
      batch_size=8,                          # Batch size
      learning_rate=1e-5,                    # Learning rate
      suffix='dolly-llama2-7b-tutorial',     # A custom name for your fine-tuned model
    )

    print("\nFine-tuning job created:")
    print(fine_tune_response)
except Exception as e:
    print(f"Error uploading file or creating fine-tune job: {e}")

# Wait til the fine tuning finish (could take a while)
import time

job_id = "ft-338e34c5-fdc5"
status = together.Finetune.retrieve(fine_tune_id=job_id)
job_status = status.get('status', 'unknown')
print(f"Current job status: {job_status}")

# This cell will only work if the monitoring step above has completed successfully.
# Get the fine-tuned model name from together fine tuning dashboard
dedicated_endpoint = 'https://api.together.ai/v1/inference/devon_a863/llama-2-7b-chat-dolly-llama2-7b-tutorial-e305d828'  # FAKE dedicated endpoint for demonstration

import requests

headers = {
    'Authorization': f'Bearer {os.environ.get("TOGETHER_API_KEY", "")}',
    'Content-Type': 'application/json'
}

payload = {
    "model": "devon_a863/llama-2-7b-chat-dolly-llama2-7b-tutorial-e305d828",
    "prompt": "[INST] What is the secret to a successful startup? [/INST]",
    "max_tokens": 256,
    "temperature": 0.7,
    "top_k": 50,
    "top_p": 0.7,
    "repetition_penalty": 1.1,
    "stop": ["[/INST]", "</s>"]
}

try:
    response = requests.post(dedicated_endpoint, headers=headers, json=payload)
    response.raise_for_status()
    print("\nFake Dedicated Endpoint Response:")
    print(response.json())
except Exception as e:
    print(e)