Request for BEHAVIOR-1K post-trained G0.5 checkpoint / training recipe

[didida-zhang]

Hi Galaxea Team,

I am studying the G0.5 paper and trying to reproduce the BEHAVIOR-1K experiments.

I noticed that the paper reports BEHAVIOR-1K results for:

• G0.5 (1 epoch post-training): 0.2904
• G0.5 (4 epoch post-training): 0.3136

Could you please clarify:

1. Are the BEHAVIOR-1K post-trained checkpoints released or planned to be released?

2. If not, is there any plan to release:

   • the fine-tuning script/config,
   • the processed BEHAVIOR-1K training format,
   • or the final checkpoint?

3. For reproducing the BEHAVIOR experiments:

   • Is the post-training simply supervised fine-tuning with the same next-token prediction objective?
   • Were CoT tokens enabled during BEHAVIOR post-training?
   • Was the visual memory module enabled during evaluation?

4. Are there any embodiment-specific modifications for R1-Pro in the BEHAVIOR setup?

Thanks!