Tony Wu

My hands are full at the moment, so I'll have to pass sorry @ariG23498 !
But I'll be more than happy to further discuss VLM-related research and training tricks on X (I think we already follow each other anyway 😉).

Thank you for the great work! Can I suggest a few things? 🤗

1. Imho, the plots (that look great btw) would be less confusing if your original dataset was shuffled, e.g. the figure is "Greedy packing" doesn't look like what you'd get in practice.
2. On top of balancing the number of images in Stage 5, balancing the number of examples would also help training. For example—looking at your figure in Stage 4—there is a high difference in the number of examples per sequence (what you refer to as batch) btw sequence 1 and the last sequence. NVIDIA's EAGLE 2 paper (Li et al., 2025, https://arxiv.org/abs/2501.14818) shows that using a balanced version of knapsack helps training! (see figures 9 and 10). Just thought it'd be nice to share this technique with the community!

But again, really nice work on this blog post and on Picotron guys!