about dataset

[chenxiahan]

Thanks for your excellent work！
Could you please share the additional jsonl for QA or the script for generation in the stage1 pretraining.

[SII-Whaleice]

thx!
I have update readme and add the qa_jsonl in hf https://huggingface.co/SII-Whaleice/SGDrive/tree/main

[chenxiahan]

thx! I have update readme and add the qa_jsonl in hf https://huggingface.co/SII-Whaleice/SGDrive/tree/main

Thank you very much！

Could you provide more details about training for all three stages, such as batch_size, per_device_batch_size, epoch and so on.
I found the setting in the code

SGDrive/internvl_chat/shell/internvl3.0/2nd_finetune/internvl3_2b_dynamic_res_2nd_finetune_wm.sh

Line 49 in 7cb55b1

--num_train_epochs 1 \

different with that in paper.

Thanks again.

[SII-Whaleice]

I recall that the configuration at the time used 32 H20 GPUs, with a total batch size of 32, a per-device batch size of 1, and 3 epochs.

[chenxiahan]

I recall that the configuration at the time used 32 H20 GPUs, with a total batch size of 32, a per-device batch size of 1, and 3 epochs.

Thank you. I will try it.

By the way, what is the settings in SGDrive-IL and SGDrive-RL?