I'd like to ask about methods to force the model to generate the <SEG> token.

[minjung98]

I was impressed by your excellent work. I wanted to ask about something - to accomplish the task you proposed, it seems that the token must appear in the output sequence. I am working on a similar task, but when I tried, even though I included the token in the training data, the model doesn't seem to include it in its responses. I'd like to ask if you used any special methods to force the generation of the token, or if you were able to achieve such results simply through finetuning with the training data.
Thank you for reading my question.

[2kxx]

Hi, minjung98)! I meet the same question. Do you figure out it?

[minjung98]

I figured it out referring https://github.com/mbzuai-oryx/groundingLMM
It it almost same code, but it was work for me.

[2kxx]

thanks

…

------------------ Original ------------------ From: minjung98 ***@***.***&gt; Date: Thu,Mar 6,2025 7:06 PM To: dvlab-research/LISA ***@***.***&gt; Cc: 2kxx ***@***.***&gt;, Comment ***@***.***&gt; Subject: Re: [dvlab-research/LISA] I'd like to ask about methods to force themodel to generate the <SEG&gt; token. (Issue #170) I figured it out referring https://github.com/mbzuai-oryx/groundingLMM It it almost same code, but it was work for me. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt; minjung98 left a comment (dvlab-research/LISA#170) I figured it out referring https://github.com/mbzuai-oryx/groundingLMM It it almost same code, but it was work for me. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***&gt;