Datasets:

---

IntelLabs

/

Intel_Robotic_Welding_Multimodal_Dataset

Like 62

Follow

Intel Labs 97

@amr-lopezjos

Hi - adding to this thread, because the access queue for this dataset appears to be unmonitored rather than simply slow.

In discussion #1, three separate people have requested access since April 2026 and none received a reply. My own request has been pending since 31 July 2026 (account: ssw08). The contact address on the dataset card, jose.a.lopez@intel.com, was reported in this thread as unreachable, and @amr-lopezjos is the only contact listed but is not a member of the IntelLabs organization here - so I suspect the approval notifications are not reaching anyone who can act on them.

Tagging three people who plausibly can:

@tarakt - co-author of arXiv:2409.02290 and an IntelLabs org member

@vagheshpatel - published Intel/qwen3.5-2b-vlm-weld-explainability-lora, trained on this dataset, last month

@michaelbeale-il - Intel / Intel Labs org maintainer

For context: I am a trainee at the AI Academy, a national AI talent development program funded by the Ministry of Science and ICT (MSIT) of the Republic of Korea. My final capstone project is a proof-of-concept for weld defect detection, due December 2026. It is strictly non-commercial and for educational assessment, not a deployed inspection system. I will comply fully with the Intel Research Use License Agreement and am happy to provide written confirmation of enrollment from the program office.

Could someone review the pending queue, or point us to the current owner of the repository?

To be clear about scope: I am asking for the multimodal data, not the images alone. The audio matters for my project specifically - the paper reports audio outperforming video on porosity (0.995 vs 0.987 AUC), overlap (0.974 vs 0.852) and crater cracks (0.945 vs 0.844), and porosity is one of my target defect classes. If releasing the full 39.9 GB archive to external requesters is not practical, a reduced subset of the FLAC audio together with the post-weld images and their labels would already be enormously useful.

One related question, in case it is easier to answer than the access request: the dataset card lists the audio as 1 channel at 16 kHz, while the paper describes two high-bandwidth microphones recording at 192 kHz. Is the published version downsampled relative to the data used in the paper? Knowing that would help me set realistic expectations before building on the audio branch.

Thank you for making this work public.

Hi @tarakt @vagheshpatel @michaelbeale-il ,

Following up on my comment above. My access request was approved on Aug 18, shortly after I posted here, and I wanted to say thank you to whoever took the time to review it. I also appreciate your patience with being tagged.

We've been working with the data for about six weeks now on our capstone project, and it has been really useful. The labels and the per-sample sensor data made it easy to get started.

I'd like to ask about something, and please feel free to say no. Do you have any separate data for validating models, something collected outside the setup used for this release? For example, welds from a different cell, a different day, or an actual production line. If so, would it be possible to share it, even a small portion?

The reason I'm asking: our model does well on held-out splits of this dataset, but its performance dropped sharply when we tried it on welding images from other sources. We can't yet tell how much of that comes from our model and how much comes from differences in camera and lighting. Some data from similar sensors under different conditions would help us understand this.

If nothing like that is available, or it can't be shared outside Intel, I completely understand. Even knowing that would help us plan. We would of course follow the same license terms. Also, if there's an email address we could use for questions like this, I'd be grateful if you could share it.

Thanks again for releasing this dataset,
Seongwoo Shin (ssw08)