SAM3 Fine-Tuning: Early Stopping Fails, Training Won't Stop Even After Manual Cancel - Monthly Credits Exhausted + 4x Overconsumption, No Response from X/Mail Support

🤝 Community Help
1

Hi Roboflow Team and Community,

Urgent issue with SAM3 fine-tuning: I enabled early stopping (patience: X epochs), but it never kicked in. Manual stop via dashboard left it hanging in “Stopping…” and credits kept draining!

Outcome: Entire monthly quota wiped out, plus 4x overconsumption! (Started with full credits, expected minimal use, now deeply negative.)

I’ve tried contacting:

  • Email support
  • X (@roboflow mention and chat)
    No replies after several days. This is blocking my project badly.

CleanShot20251201-23-41-16
CleanShot20251201-23-41-161546Ă—1546 230 KB

Requests:

  1. Full refund for overcharged credits.
  2. Fix for early stopping bug (similar issues reported?).
  3. Prevention tips for future runs.

Attached screenshots: Credit balance before/after (showing overage), dashboard during stop attempt, early stopping settings.

Need help ASAP! Thanks,

2

Hello Tao.
I’m Tony, from Roboflow engineering team.
Didn’t look at your case yet. Just wanna tell you that I hear you and will start investigating.
Be back soon.

3

Hello Tao.

  • I refunded 34 credits regarding your failed training
  • bug confirmed. mid-train stop is not working properly for SAM3. We’re on it.
  • SAM3 finetuning is pretty resource-intensive. For faster training, I recommend rf-detr-seg which is also pretty good.

If you want to try out SAM3, I recommend starting with smaller datasets (~300-500 images) to get a feel for how the model behaves.
Important note: inference for finetuned SAM3 models doesn’t work in our hosted inference solution yet. But you can spin up a dedicated deployment or run our inference server yourlself (needs GPU-powered hardware)