Personal UCL Infra

The lab PCs boot into Linux by default and can potentially be accessed remotely. However please note that the PCs regularly reboot on Monday and Thursday evenings between 7.30pm and midnight. These PCs are also used by others and may be rebooted at any time. If you wish to run a long compute job on a lab PC a good option would be to run the job overnight, excluding Monday and Thursday evenings, or at the weekend.

seongcho

oR12RtZihcGJ

kamzi	1080
canada	3090ti
shelduck	3090ti
breeze	1080

4673 for lab password

request

Grants:

Leave this blank unless you have specific research grants tied to this project. You can update it later if needed.

Research Themes:

Based on REF categories, your research would likely fall under Artificial Intelligence and Informatics or Computer Science as it is focused on advancing AI methodologies.

Types of resource needed:

You will likely need access to:

At least one GPGPU for neural network operations (since you’re working with LLMs).

Possibly medium-sized MPI jobs (80-320 cores) for large-scale vector manipulation and benchmarking.

Extra technical requirements:

No unique extra requirements seem necessary based on the project description, unless you need specialized hardware like certain types of GPUs or software tools specific to vector extraction.

Project Name:

"Steering AI through Activation Engineering for Enhanced Reasoning"

Work description:

I will be applying Activation Engineering to improve the reasoning abilities of Large Language Models (LLMs). This involves extracting and manipulating internal feature vectors from models like GPT-4 to directly control AI behavior. The aim is to enhance AI performance in reasoning and question-answering tasks compared to traditional prompt-based control methods. The experiments will be conducted using Gemma Scope to evaluate the effectiveness of Steering Vectors on various benchmarks.

Dependencies:

PyTorch (v2.x) for model manipulation

Hugging Face Transformers (latest) for working with LLMs

CUDA (v11.x) for GPU acceleration

Gemma Scope (specific libraries associated with its usage)

Dependencies description:

The software tools will be used for manipulating the internal states of LLMs and conducting performance benchmarks. Specific versions of PyTorch and CUDA are required for neural network computation and GPU support.