Thinking mode not working correct
Hi! Thank you for the work done!
Maybe I'm doing something wrong, but when I use R1_SYSTEM_PROMPT, the model does not produce thinking content. There is no content separation method inside the chat, so I expected to see thinking content and just content in response. I used cascade RL trained models (InternVL3.5-241B-A28B and InternVL3.5-30B-A3B, tried MPO-only training versions) for images and videos. Maybe thinking is not mandatory, as it was in Qwen 3 when using the think mode, then could you tell me which prompt to use to regularly cause the model to think? I will be grateful for any answer.
Using R1_SYSTEM_PROMPT:
After that, an example in which thinking content was expected, but did not appear:
Thank you for your interest in our paper. This issue may arise because our model has not been trained on captioning tasks in the thinking mode. To ensure that thinking mode is reliably triggered, you may consider replacing the following code segment in modeling_internvl_chat.py.
Specifically, replace
history = [] if history is None else history
for (old_question, old_answer) in history:
template.append_message(template.roles[0], old_question)
template.append_message(template.roles[1], old_answer)
template.append_message(template.roles[0], question)
template.append_message(template.roles[1], None)
query = template.get_prompt()
with
history = [] if history is None else history
for (old_question, old_answer) in history:
template.append_message(template.roles[0], old_question)
template.append_message(template.roles[1], old_answer)
template.append_message(template.roles[0], question)
template.append_message(template.roles[1], '<think>')
query = template.get_prompt()
if query.endswith(template.sep):
query = query[:-len(template.sep)]