Question About New Released InternVLA-A1-3B Checkpoint

[MaokeAI]

Hi, thanks for releasing InternRobotics/InternVLA-A1-3B.

I have a question about how to interpret the training information saved in the released model files.

From the released model directory, I found that train_config.json contains entries such as:

• resume = true
• policy.pretrained_path = outputs/.../checkpoints/01310000/pretrained_model
• steps = 25000000
• batch_size = 28
• policy.scheduler_decay_steps = 25000000

Because of this, I am not sure how to correctly understand the actual training stage of the released checkpoint.

My current interpretation is:

• the released model may have been resumed from a previous pretraining checkpoint at around 1,310,000 steps
• steps = 25000000 may describe the planned maximum training budget for that resumed run, rather than the actual completed step count of the released model
• batch_size = 28 in train_config.json may describe a local/per-process setting from that resumed run, but I am not sure how it relates to the batch size reported in the paper

Could you please clarify:

1. What was the actual training step of the released InternVLA-A1-3B model when it was exported?
2. Is the 01310000 checkpoint in policy.pretrained_path really a 1.31M-step pretraining checkpoint that was used to initialize the released model?
3. Does steps = 25000000 mean the total planned budget of the resumed training run, or the real completed step count?
4. In the paper, the reported global batch size seems to be 512, but in train_config.json I see batch_size = 28. Could you clarify how these numbers relate to each other? For example, is 28 a per-GPU/per-process batch size from a particular resumed run, while 512 is the effective global batch size used in the main training setup?
5. If possible, could you describe the training lineage of the released model more concretely, e.g. “started from checkpoint X at step 1.31M, then continued training on datasets A/B/C, and the released weights correspond to approximately step Y”?

I am asking because I want to compare the released model fairly with another local checkpoint and avoid misinterpreting the training stage or batch-size setting.

Thanks.