https://github.com/daveshap/raspberry_experiments
Keeping my personal experiments separate from the main repo
https://github.com/daveshap/raspberry_experiments
Last synced: about 1 year ago
JSON representation
Keeping my personal experiments separate from the main repo
- Host: GitHub
- URL: https://github.com/daveshap/raspberry_experiments
- Owner: daveshap
- License: mit
- Created: 2024-09-14T08:11:37.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2025-02-11T11:30:15.000Z (over 1 year ago)
- Last Synced: 2025-04-24T03:46:09.276Z (about 1 year ago)
- Language: Python
- Size: 77.1 KB
- Stars: 65
- Watchers: 10
- Forks: 6
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Raspberry Experiments
Mission: Create an open source finetuning dataset that allows anyone to train AI models to engage in "Test Time Comput" or chain of thought reasoning.
## Process
For starters, OpenAI GPT-4o only needs 50 to 100 samples to get started. We can demonstrate success by simply improving what the base model does. For instance, if the base GPT-4o scores 25% on a given benchmark, but we can boost that to 32% with a simple finetuning job, that is a solid proof of concept. So our goal is simply to create a statistically significant lift in performance.
1. We need some high quality queries that would be realistic from a user perspective. These can be synthesized easily enough.
2. We need some high quality chains of thought (reasoning) as the answer. Ideally they would be correct answers, but most importantly we just need to train the model to use reasoning.