To systematically assess the impact of RLHF on the truthworthyness of GPT’s answers it is mandatory to have access to the underlying pretrained-only models. Do I understand it correctly that all models that can be used over the API are fine-tuned by RLHF? So this kind of research is not possible? Which workarounds are at hand?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| Prime Zero Shot Learners towards Factfulness - GPT3 vs. GPT-J-6B | 3 | 623 | June 23, 2021 | |
| ChatGPT 3.5 was not trained with RLHF | 1 | 2262 | November 20, 2023 | |
| GPT-4 System Card by OpenAI - March 15, 2023 | 3 | 5476 | December 17, 2023 | |
| Just got access to GPT-4 but it responds like 3.5 | 13 | 8216 | July 8, 2023 | |
| Data cutoff date in 1106 models | 6 | 9696 | April 12, 2024 |