Miguelito De Guzman PRO
migueldeguzmandev
AI & ML interests
None yet
Organizations
These models represent the 10 training RLLM checkpoints/ layers intended to improve GPT2XL's alignment to an ethical persona.
in-progress
In-progress, training started April 22, 2024; See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
A new build, see: visual map for more details. https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
This has been the best build alignment (so far), for more info: https://www.lesswrong.com/posts/x5ySDLEsJdtdmR7nX/rllmv10-experiment
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
This is a collection designed to present the ten RLLM steps/ training runs intended to improve Phi-1.5's outputs towards coherence and politeness.
(swapped truth dataset to Q&A); failed at jailbreaks though..
RLLM trained models with strong robustness to jailbreaks
Experiment in progress: reverse/cure/reduce Paperclip PeterTodd persona (PPT) using RLLMv3 datasets...
Version 1 of Reinforcement Learning using Layered Morphology (RLLM)
first two training steps, training halted; See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
Spin-off experiments eg. what will happen if the v10/layer10's weight decay got increased or decreased....so yeah all these attempts failed..
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
Training checkpoints for RLLMv7 experiment.
This is a collection designed to present the 10 RLLM steps/ training runs intended to improve Falcon-RW-1B's outputs towards coherence and politeness.
A spinoff projet - chaining RLLMv1 to RLLMv3.....(exploratory, Since it's possible to chain PaperclipTodd and RLLMv3)
RLLM trained models with strong robustness to jailbreaks
These models represent the 10 training RLLM checkpoints/ layers intended to improve GPT2XL's alignment to an ethical persona.
Experiment in progress: reverse/cure/reduce Paperclip PeterTodd persona (PPT) using RLLMv3 datasets...
Version 1 of Reinforcement Learning using Layered Morphology (RLLM)
in-progress
In-progress, training started April 22, 2024; See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
first two training steps, training halted; See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
A new build, see: visual map for more details. https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
Spin-off experiments eg. what will happen if the v10/layer10's weight decay got increased or decreased....so yeah all these attempts failed..
This has been the best build alignment (so far), for more info: https://www.lesswrong.com/posts/x5ySDLEsJdtdmR7nX/rllmv10-experiment
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
See RLLM Visual Map for details, https://whimsical.com/rllm-visual-map-QQvFHNr6aVDdXRUnyb5NCu
Training checkpoints for RLLMv7 experiment.
This is a collection designed to present the ten RLLM steps/ training runs intended to improve Phi-1.5's outputs towards coherence and politeness.
This is a collection designed to present the 10 RLLM steps/ training runs intended to improve Falcon-RW-1B's outputs towards coherence and politeness.
(swapped truth dataset to Q&A); failed at jailbreaks though..
A spinoff projet - chaining RLLMv1 to RLLMv3.....(exploratory, Since it's possible to chain PaperclipTodd and RLLMv3)