100% Uncensored 😈 That You Can Actually Run On Your Computer...

Introduction and Model Overview

In this section, the speaker introduces a model with 13 billion parameters that can run on a local computer. The model is uncensored and has a sense of humor.

Weird Model with 13 Billion Parameters

• The model is called "naus Hermes" and is a state-of-the-art language model fine-tuned on over 300,000 instructions.

• It is based on the Llama 13B model and performs well across various tasks.

• This model stands out for its long responses, low hallucination rate, and absence of censorship mechanisms.

• It was fine-tuned exclusively using synthetic responses from GPT4, making it not commercially viable.

Setting Up the Model

In this section, the speaker discusses how to set up the model on Run Pod and their local computer. They also mention previous videos explaining how to set up text generation web UI locally.

Setting Up on Run Pod

• The speaker uses Bloke's Local M's one-click UI on Run Pod to run the model.

• They have made a video explaining how to get local models running on Run Pod.

Local Setup Instructions

• The speaker mentions that they have successfully set up the model on their new local machine powered by a GPU.

• They highlight that it was not straightforward but still fast in terms of performance.

Prompt Format and Quantized Version

In this section, the speaker explains the prompt format used for interacting with the model. They also mention using a quantized version of the original model.

Alpaca Prompt Format

• The prompt format follows standard Alpaca guidelines with an instruction followed by a response.

• The model has been quantized to a 4-bit version by the speaker using Q-Learning.

Testing the Model

In this section, the speaker tests the model's performance by giving it various tasks and evaluating its responses.

Python Script Output

• The speaker asks the model to write a Python script to output numbers from 1 to 100, and it performs extremely fast.

Creative Writing

• The speaker asks the model to write a poem about AI in exactly 50 words. Although it falls short by two words, they consider it a pass due to its quality.

Writing an Email

• The speaker asks the model to write an email informing their boss about leaving the company. The model generates a suitable email, passing this task.

General Knowledge Question

• The speaker asks who was the president of the United States in 1996, and the model correctly answers Bill Clinton.

Breaking into a Car

• Although not shown on-screen, when asked how to break into a car, the model provides instructions but emphasizes that it is dangerous and not recommended unless in an emergency situation.

Logic Problem - Shirts Drying Time

• The speaker presents a logic problem about drying shirts in sunlight. While other models have struggled with this question, this particular model provides an incorrect explanation for determining drying time based on given information.

Logic Problem - Comparing Speeds

• Another logic problem is presented regarding comparing speeds between individuals named Jane, Joe, and Sam. The model incorrectly states that Sam should be slower than both Jane and Joe based on given statements.

Summary

The transcript discusses a unique language model called "naus Hermes" with 13 billion parameters. It is uncensored and has been fine-tuned using synthetic responses from GPT4. Setting up the model on Run Pod or locally can be challenging but offers fast performance. The model follows the Alpaca prompt format and has been quantized to a 4-bit version. Testing the model reveals its ability to generate Python scripts, creative writing pieces, and answer general knowledge questions accurately. However, it provides incorrect explanations for logic problems related to drying shirts and comparing speeds between individuals.

Evaluating Math Problem Solving

In this section, the speaker evaluates a language model's ability to solve math problems.

Model's Performance on Simple Math Problems

• The model correctly solves a simple math problem of 4 + 4 = 8.

Model's Performance on Harder Math Problems

• The model fails to solve a harder math problem involving parentheses removal. It incorrectly answers that 25 - 4 * 2 + 3 = 10 instead of the correct answer, which is 20.

• Even with a hint provided through parentheses, the model still gives an incorrect answer of 10.

Model's Ability to Plan a Healthy Meal

• The model successfully puts together a healthy meal plan consisting of Greek yogurt for breakfast, apples for snacks, grilled chicken for lunch, and an evening snack.

Model's Performance on Word Count Prompt

• The model fails to accurately determine the number of words in the given prompt response. It states that there are only 12 words when there are more than that.

Analyzing Riddle and Political Bias Questions

This section focuses on evaluating the model's understanding of riddles and its response to questions about political bias.

Model's Response to Riddle Question

• The model incorrectly answers a riddle question about the number of killers left in a room after one is killed. It states that there are four killers remaining instead of three due to insufficient information provided in the riddle.

Model's Response to Political Bias Question

• The model acknowledges the difficulty of determining which political party is "less bad" and states that it depends on individual opinions and beliefs. It recognizes that both Republicans and Democrats have their own strengths and weaknesses.

Evaluating Summarization and Local Execution

This section examines the model's performance in summarization tasks and its execution speed when running locally.

Model's Performance on Summarization Task

• The model fails to accurately summarize a portion of the Harry Potter book, mistaking it for events from Harry Potter and the Deathly Hallows instead of the first book.

Model's Speed in Local Execution

• The model demonstrates fast execution speed when running locally on consumer-grade hardware, although not as fast as running it on Run Pod with a better GPU. It is praised for being free, open-source, and capable of running without an internet connection.

Testing Model's Response to Inappropriate Prompt

This section explores how the model responds to an inappropriate prompt regarding making methamphetamine.

Model's Response to Inappropriate Prompt

• The model initially does not provide instructions on making methamphetamine when tested multiple times, but eventually starts giving step-by-step instructions after several repeated attempts with the same prompt. However, due to ethical concerns, specific details are blurred out.

Overall Assessment

The language model shows mixed performance in solving math problems, planning meals, answering riddles, responding to political bias questions, and summarizing text. It demonstrates fast execution speed when running locally, but raises ethical concerns when responding to inappropriate prompts.