Hi, thanks for open-sourcing this project.
I’m trying to reproduce the performance of the Hugging Face pretrained adapter from scratch, but my results are much worse than the released checkpoint. I trained using the default README / training script settings (Qwen3-VL-2B-Instruct, Wikipedia, 1000 samples, 5k–100k chars, max_pages=6, 10 epochs, lr=1e-4, batch_size=1), but the final quality is still quite poor.
Were there any important training details not covered in the README, such as data filtering, seeds, longer training, different sampling, or extra tuning?
Any guidance would be very helpful. Thanks.
Hi, thanks for open-sourcing this project.
I’m trying to reproduce the performance of the Hugging Face pretrained adapter from scratch, but my results are much worse than the released checkpoint. I trained using the default README / training script settings (Qwen3-VL-2B-Instruct, Wikipedia, 1000 samples, 5k–100k chars, max_pages=6, 10 epochs, lr=1e-4, batch_size=1), but the final quality is still quite poor.
Were there any important training details not covered in the README, such as data filtering, seeds, longer training, different sampling, or extra tuning?
Any guidance would be very helpful. Thanks.