Falcon 2What that means
A raw pretrained model continues text. Ask it a question and it may answer, or it may produce more questions in the same style, because both are plausible continuations of the text you supplied. That is not a bug, it is the model working as trained.
The one format TII does publish
The VLM checkpoint was built from a chat-finetuned Falcon 2 11B and its model card shows the format it expects: User:<image> followed by the instruction, then Falcon: to open the reply. If you want conversational behaviour from the Falcon 2 family out of the box, that is the checkpoint with a documented instruction format.
Options for the text model
- Fine-tune on instruction data, which is what TII recommends for serious use.
- Use few-shot prompting with explicit, consistent formatting you control.
- Start from the VLM backbone if you do not need vision but do want instruction following.
- Wrap the model behind your own template layer and keep it versioned: an undocumented prompt format is a hidden dependency.
Related pages
Sources
falcon2.lol is an independent third-party site. It is not affiliated with, endorsed by, sponsored by or operated by the Technology Innovation Institute, and it does not speak for TII. Every number on this page is attributed to the source listed above.