Final Submission question

PAN
1

Sort of a few questions.

#1 how do you actually make the final submission?

I’ve ran
tira-run `

–input-dataset generative-ai-authorship-verification-panclef-2025/pan25-generative-ai-detection-smoke-test-20250428-training --image pan-2025-ben-v2
–command ‘python main.py --input_file $inputDataset/dataset.jsonl --output_dir $outputDir’ `
–push true
and pushed - im assuming that the software shows up on the tira site and i make a submission? When i check the docker submission instructions page (TIRA) and click next i dont go to page 3 it sends me back

#2 when running the command above i tried to push, but got a timeout error - im trying to run an offline version of Roberta Large and my image is about 7gb - is there a size limit? How do i get around this - i was able to get a prediction.jsonl on the test set

#3 i tried to use gpu when running the local test
device = torch.device(“cuda” if torch.cuda.is_available() and args.use_gpu else “cpu”)
but from the output log it looks like gpu is not utilized

2

Hi, can you invite me to the repository (my account is mam10eks on github)? I can have a look and help to finalize the submission.

If I recall correctly, the size limit is 15 GB per layer, so 7 GB should be fine.

Is the model publicly available on hugging face? Then, we could also mount it to the container, I can also help to configure this in the repo.

It might be that cuda is not installed in the docker image, then torch.cuda.is_available() returns false even when the host has a GPU.

Best regards,

Maik

3

Thanks so much maik!

I’ve added you to the repo im trying to submit please have a look when you can.

the model is not publicly on huggingface i’ve just tried to upload a docker submission

I will look into adding cuda into the docker image - i currently have a run being pushed thats been processing for 50 minutes - fingers crossed!

4

Perfect, for the cuda thing, the easiest option is to start from an image that already has it installed (otherwise installing cuda easily gets difficult), e.g., for the generative-authorship-verification baseline, we used:

FROM nvcr.io/nvidia/cuda:12.8.0-cudnn-devel-ubuntu24.04

I will have a look at the repo.

Best regards,

Maik