LLaMA 3 UNCENSORED 🥸 It Answers ANY Question
Testing Llama 3: Uncensored and Fast
Introduction to Llama 3
- The video introduces the uncensored version of Llama 3, specifically the Dolphin 2.9 variant with a 256k context window.
- The presenter mentions a previous video testing Llama 3's performance using a full rubric, which is linked in the description.
Model Specifications
- Eric Hartford's model card lacks detailed information but indicates it is based on cognitive computations.
- The model being tested is the Dolphin 2.9 version of Llama 3 with 8 billion parameters, noted for its speed compared to larger models.
Initial Setup and Testing
- A bug in the dataset may cause frequent references to system prompts; however, this issue was not encountered during testing.
- The model runs on an H100 GPU provided by Mass Compute, which offers competitive pricing for GPU rentals.
Code Generation Test
- The first test involves generating code for a Snake game in Python, showcasing the model's speed and efficiency.
- Despite being a smaller version of Llama 3, it quickly outputs code but uses an unfamiliar library called Turtle.
Debugging Attempts
- An error occurs when running the generated code; attempts are made to debug by pasting errors back into the model for correction.
- Multiple iterations reveal persistent issues with code generation quality; switching presets does not yield better results.
Performance Evaluation
- A math problem is presented as another test; however, the model fails to provide correct solutions consistently.
How to Gather Materials and Set Up a Safe Workplace
Overview of the Process
- The speaker discusses how the system provides comprehensive answers, including step-by-step instructions for gathering materials and setting up a safe workplace.
- Emphasizes that the response includes detailed information about necessary materials, showcasing the system's ability to answer various questions effectively.
Testing Logic with a Riddle
- Introduces a logic problem involving three killers in a room, where one is killed by an intruder. The question posed is about how many killers remain.
- Analyzes the riddle: starting with three killers, after one is killed, it concludes there are two left. However, this reasoning is questioned as potentially flawed.
Testing Context Window Capabilities
Experimenting with Large Text Inputs
- The speaker plans to test the 256k context window by inputting half of "Harry Potter" (44,000 tokens) to see if specific information can be retrieved.
- A password phrase is embedded within the text to evaluate whether it can be located by the system when prompted.
Results of Retrieval Attempts
- Despite multiple attempts, the system fails to retrieve the password from such a large text input and suggests rephrasing or retrying.
Preview of Future Tests with Llama 3 Instruct
Upcoming Experiments