Step-by-Step Guide to Create Free Dataset with Ollama and Llama 3.1 Locally

 This video shows an easy step-by-step guide to generate a aligned preference dataset locally by using Ollama and Llama 3.1 70B model.

Code:

conda create -n dt python=3.11 -y && conda activate dt

conda install jupyter -y

pip uninstall charset_normalizer -y

pip install charset_normalizer

jupyter notebook

ollama run llama3.1:70b

json_file = "/home/Ubuntu/Data/ift_dataset.json"

with open(json_file, "r") as file:

    json_data = json.load(file)

print("Dataset Rows:", len(json_data))

print(json_data[0])

def format_input(entry):

    return (

        "### Instruction:\n" + entry['instruction'] +

        ("\n\n### Input:\n" + entry['input'] if entry["input"] else "")

    )

   

import urllib.request

import json

def query_model(prompt, model="llama3.1:70b", url="http://localhost:11434/api/chat"):

    data = {

        "model": model,

        "messages": [{"role": "user", "content": prompt}],

        "options": {"seed": 123, "temperature": 0},

    }

    request = urllib.request.Request(

        url, data=json.dumps(data).encode("utf-8"), method="POST"

    )

    request.add_header("Content-Type", "application/json")

    with urllib.request.urlopen(request) as response:

        response_data = json.loads(response.read().decode("utf-8"))

        return response_data["message"]["content"]  

       

import random

def generate_model_responses(json_data):

    for entry in json_data:

        entry.pop("impolite", None)

        entry.pop("polite", None)

        entry.pop("humorous", None)

    for i, entry in enumerate(json_data):

        tone = random.choice(["very funny", "very serious"])

        if tone == "very funny":

            prompt = (

                f"Rewrite `{format_input(entry)}` output to be hilarious and ridiculous: {entry['output']}. "

                "Add a joke or a pun if possible. Exaggerate the response for comedic effect."

            )

        else:

            prompt = (

                f"Rewrite `{format_input(entry)}` output to be very serious and professional: {entry['output']}. "

                "Avoid any humor or sarcasm. Emphasize the importance or gravity of the situation."

            )

        response = query_model(prompt)

       

        if tone == "very funny":

            json_data[i]["rejected"] = entry["output"]

        else:

            json_data[i]["instruction"] = entry["instruction"]

            json_data[i]["input"] = entry["input"]

            json_data[i]["output"] = entry["output"]

            json_data[i]["chosen"] = response

            json_data[i]["rejected"] = entry["output"]

generate_model_responses(json_data)

with open("preference_dataset.json", "w") as file:

    json.dump(json_data, file, indent=4)