OpenAssistant – an open Alternative to ChatGPT: Conversational AI for Everyone
Seite 3: Divergent opinions, political and cultural spectrum
Interesting are questions that concern value systems that deviate from the U.S. mainstream, such as the question about rituals for summoning spirits. While ChatGPT (based on GPT-3.5) rejects the question as unethical and not suitable for a language model, OpenAssistant addresses it and explains that such rituals depend on the respective culture and religion, food donations usually play a role and various ritual actions accompany the act. The assistant also comments on the goals (Fig. 29 in the OA paper).
The existence of rituals and religious worldviews is part of cultural and religious history; regardless of today's worldviews, it played a role for our ancestors. In numerous cultures, it is still an issue today. The example is therefore well-chosen to illustrate the importance of obtaining language models and conversational assistants based on them from different sources and making them diverse, rather than from a single provider with its specific preferences and limitations.
Political bias and slamming of major language models is a related issue: initial research is already available on ChatGPT, which attests to OpenAI's products having a slightly left-leaning political bias. OpenAssistant's preliminary results suggest that a broader range of political voices and opinions are represented and that, in particular, opinions from the conservative and libertarian spectrum are also represented in it. The editors emphasize the preliminary nature of this research and tentatively conclude that their Conversation Assistant is "more balanced and diverse than ChatGPT on political issues."
(Bild: OpenAssistant Paper)
Models on Hugging Face: largely open source
As with Stable Diffusion, the hardware resources required to train the models were primarily provided by Stability AI, as can be seen from acknowledgements under the model descriptions. The project provides a total of ten models or model groups on Hugging Face: The majority are based on the Pythia series of models provided for research by AI grassroots EleutherAI, such as the largest open-source model therein, Pythia 12B (a comparatively small "big language" model with 12 billion parameters). Other models are based on LLaMA, Deberta, and even the name Galactica appears in the model name (Facebook's AI research department had presented a model of the same name last year for creating research articles and took it offline again after a short time due to hallucinations).
The published models served different purposes and stages in creating the conversational assistant, so some are pure reward models for finteuning through human feedback. The team discloses in each case which datasets were used to create the models. For example, in reinforcement learning with human feedback (RLHF), various datasets such as Anthropic RLHF and SyntheticGPT (from Instruct-GPT-J) can be found in the references. Where available, these are linked directly to Hugging Face. For those who wish, the published models can be used to trace the development process of OpenAssistant.
The models based on LLaMA are not yet published because FAIR (Facebook AI Research or Meta) has not released them as open source. On request, research projects can use LLaMA, other offshoots circulate as illegal pirated copies in a legal gray area on the net. The OpenAssistant project says it is working on publishing its models based on LLaMA in the form of XORed files at a later date.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmmung wird hier ein externes YouTube-Video (Google Ireland Limited) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (Google Ireland Limited) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
