Request to update experiment scripts, code and data

[lerogo]

Thank you for your excellent work.
We now need to reproduce your work and kindly request that you quickly update the experiment scripts and data.

Thank you again for your work and looking forward to your reply.

[Osilly]

Thank you for your interest in our work. The dataset is expected to be released within 1–2 weeks. Due to company policies, the code is currently undergoing internal review and approval procedures. We will make the complete source code publicly available upon completion of this process.

[Osilly]

The cold-start initialization dataset has been published (https://huggingface.co/datasets/Osilly/Vision-R1-cold).

[lzcomeon]

The cold-start initialization dataset has been published (https://huggingface.co/datasets/Osilly/Vision-R1-cold).

hello, there is no prompts like "output the thinking process in tags" in the cold-start SFT data, but the response has special tokens like "" and thinking contents.

So, I wonder if there is additional prompts like "output the thinking process in tags" are added when conduct SFT cold-start?
Thank you again for your work and looking forward to your reply.

[Osilly]

@lzcomeon Thanks for your attention to our work! In the cold-start initialization stage, the output formatting has been constrained by the Vision-R1-cold dataset. Thus, we do not need to use the additional prompt like system prompt.

[Anirudh257]

Hi @Osilly thanks for releasing the cold dataset. Can you share the checkpoints and remaining scripts?

[Osilly]

@Anirudh257 The cold-start training guide is available here. The checkpoints will be released later.

[Osilly]

The checkpoint of Vision-R1-7B has been published.

[Osilly]

RL Dataset (Vision-R1-rl), Vision-R1-CI-7B, Vision-R1-32B, Vision-R1-72B has been published.

[huaiyizhao]

I have two questions regarding the cold start data:

1. You train the model to format its output like without mentioning the format in the prompt, will this hurt the model's generalization abilities? I.e., the model will always output as its first response token.
2. In the paper, you said "we use the multimodal visual question answering (VQA) datasets, LLaVA-CoT dataset (100K) [17] and Mulberry dataset (260K) [15], to conduct Vision-R1-cold (200K).". I am confused here, does the data generation pipeline proposed in the paper are only applied on these two datasets to obtain the corresponding cot data, OR there is another VQA dataset(which I did not find on the hf link) that is constructed by the data generation pipeline?

[Osilly]

@huaiyizhao For the first question, we aim to transform a base MLLM into a reasoning-capable model without preserving its direct answer generation ability. Thus, we employ formatted training data that without system prompt to fine-tune the base model. Subsequently, we leverage the VQA data from the LLaVA-CoT and Mulberry datasets, augmenting them with CoT annotations (Please refer to here).

[peterant330]

@huaiyizhao For the first question, we aim to transform a base MLLM into a reasoning-capable model without preserving its direct answer generation ability. Thus, we employ formatted training data that without system prompt to fine-tune the base model. Subsequently, we leverage the VQA data from the LLaVA-CoT and Mulberry datasets, augmenting them with CoT annotations (Please refer to here).

Hi, thanks for your great work. I wonder during the RL phase, do you use system prompt to constrain the format of the response? I notice in the paper section 3.1, you mentioned there is system prompt, but I cannot find the complete version in the github. Also when I use Vision-R1-CI-7B without system prompt, only 20% of the response are with the correct format.

[Osilly]

@peterant330 During the RL stage, we do not use the system prompt. We find that the answer follows the expected format of <think></think><answer>Final Answer:</answer>.