Text Generation
Transformers
Safetensors
English
falcon_mamba
conversational
Inference Endpoints

Trained on ChatGPT Responses?

#9
by astrologos - opened

I told falcon-mamba its own name, but falcon-mamba told me it was not falcon-mamba. Rather, it said it was a model trained by OpenAI. Were ChatGPT responses used to train this model? If so, is this data part of RefinedWeb? If not, what’s the issue?

astrologos changed discussion status to closed
astrologos changed discussion status to open
astrologos changed discussion status to closed
astrologos changed discussion status to open
Technology Innovation Institute org

Hi @astrologos
Thanks for your message. I can confirm we did not trained the model explicitly on ChatGPT prompts. However note that the common web datasets (including instruction datasets) that are present nowadays to train and fine tune LLMs are quite contaminated with synthetic data generated from LLMs such as ChatGPT (see for example the screenshot below, taken from: https://huggingface.co/spaces/HuggingFaceFW/blogpost-fineweb-v1).
Screenshot 2024-08-18 at 12.25.43 PM.png
One technique to overcome this is to perform further model alignment by doing identification tuning so that the model "knows" he is not ChatGPT but another LLM. I suggest to stay up to date in this organization as we will release more models in the future which are better in all sense, including identification and alignment

Hi @ybelkada , many thanks for your detailed response. I have a lot of hope for SSMs. Hopefully someday we'll see constant-time Deep Koopman LLMs ;)

Sign up or log in to comment