Training Hyperparameter for Colqwen2.5-v0.2 reproduce
Hi,
I’m currently working on reproducing the fine-tuning of ColQwen2.5 and evaluating it on the Vidore benchmark. I’ve noticed that there are three different sources specifying hyperparameters for ColQwen2.5: the Hugging Face model card, the adapter_config.json file, and the training_config.yml, as well as the training script from the Colpali GitHub repository.
Could you kindly share the exact hyperparameter settings used for achieving the top results with ColQwen2.5 v0.2 on the Vidore benchmark? It would be incredibly helpful for ensuring reproducibility.
Thanks a lot!
I'm looking for a reproduction tutorial for ColQwen2.5. Could you share one?
The training_config.yaml in this model repository is the config for this model.
Make sure that batches are well shared between GPUs and not individual per GPU if you want good performance (python launch instead of default torch distributed).
Note that this model is super undertrained so you can surely easily do better