Hyphens-dashes converted to underscores for trigger words/file names

[ShaunCassidyPoster]

Win 10; Intel Core i7 930 lol; 12 GB RAM; RTX 3060 12 GB.

Description: When creating a LoRA, projects named with hyphens, e.g. gumby-v1, will have the hyphen converted to underscore in the resulting safetensors file names: gumby_v1-step00000300.safetensors. At least one follow-on effect is that the Lora Manager custom node is unable to pull the trigger word into its TriggerWord Toggle node, and perhaps other nodes that explicitly associate the trigger word with the LoRA file name.

On generation, the Lora Manager TriggerWord Toggle node produces the error:

Traceback (most recent call last):
  File "F:\ComfyUI\ComfyUI\custom_nodes\comfyui-lora-manager\py\routes\lora_routes.py", line 148, in get_lora_trigger_words
    trigger_words = await self.service.get_lora_trigger_words(lora_name)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\ComfyUI\ComfyUI\custom_nodes\comfyui-lora-manager\py\services\lora_service.py", line 260, in get_lora_trigger_words
    return civitai_data.get("trainedWords", [])
           ^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'get'

However, the LoRA apparently works. I say "apparently" because, although I use the hyphenated trigger word in the prompt, I am using Lora Manager and it may make using trigger words in the prompt unnecessary. I suppose I should know this, but I don't.

To Reproduce:

• Create a dataset and project containing a hyphen
• March through the training to create at least one checkpoint
• Note that the trigger word part of the file name now has an underscore instead of a hyphen

Expected behavior:

• I expected that part of the file name to be unchanged, since it is the basis of my dataset, prefixing every image and text file name.

Finally, despite my bug, thank you very much for making TrainFlow. LoRA training has always had a steep learning curve for a non-developer, and the time it would take to get up to speed was daunting. Seeing the likeness of my subject appear at step 600 was astounding. I hope TrainFlow catches on.

[ThetaCursed]

Thanks for your feedback and detailed description of the problem. I'll see what I can do about it.
I expected that part of the file name to be unchanged
Well, yeah, I would expect that to be the default behavior too.

[ShaunCassidyPoster]

Well, Gemini has told me (fwiw, I wouldn't recognize confabulation here): "1. Change the Trigger Word Separator... Do not use a hyphen (-) in your trigger word. Most Danbooru-based tokenizers (like the one inside Anima-Trainflow) read hyphens as separators, breaking your trigger word into three distinct tokens..." about another lora I'm trying.

Maybe the easiest "fix" is just to mention not to use dashes in the filenames or trigger word in the readme somewhere. I will re-do the first one with underscores, and it will likely go better. Despite that, the lora still works okay. It's a mysterious world.