Spaces:
Configuration error
Older models then 10/19/26
Any gemma 4 based model older then 10/19/26 should be considered a bad tune on this repository (my models, not @AesSedai or others)
If you are running into issues and would like a model updated, post here. I'll eventually get around to it... on my own time.
26B-A4B/12B too, or only 31B?
ALL Gemma 4 models.
Yeah I screwed the pooch on this one.
edit oops, i saw you just had a Wicked-Tutu-31B-Final. that looks brand new, so it's probably fixed. disregard.
v4 and final on wicked tutu are fixed, but final is really the best version released to date so use that.
To verify, All models or just the 31B's?
ALL Gemma 4 models.
Nevermind. I see that reply. I'll look over and see which ones i need to purge then and redownload.
To verify, All models or just the 31B's?
ALL gemma 4 models.
You have to setup a special chat template for training Gemma 4 models on axolotl.
Here is a direct quote:
"Why thinking breaks: Gemma 4 expects every reply to start with an empty thought block, <|channel>thought\n<channel|>. Gemma's chat template only writes it for the reply being generated, never for the replies inside a conversation. axolotl trains every reply, so every one teaches "start answering right after <|turn>model", and that kills the thought opener. The fix is a training template that puts the empty block before every reply, masked (seen, not trained). That's the setup tuners whose Gemma 4 thinking survives use. Your data doesn't need the block; the template adds it"
In addition to the above, I never trained the gate layer on previous 31B models. So with this combination, you get a broken model, essentially.
But yes the template issue affects 26BA4B, even though I did train the gate layer on 26BA4B models....
Scraping from your activity (you mentioned the last 7 months?) the models in question should likely be:
If i got any wrong let me know probably make note of them. This should make finding them easier for purging to redownload when they are ready.
Brisk-Evolution-12B
Brisk-Evolution-9B
Dark-Scarlett-v0.?-31B
Dark-Scarlett-v0.?-26B
For-Her-Darkside-12B
For-Her-Darkside-26B
For-Her-Darkside-31BFroopert-31B
G4-MeroMero-v2-31B
Gemma-4-Dark-Thoughts-31B
Gemma-4-Giftige-Blume-31B
Gemma-4-Orthia-31B
Glistening-Gem-31B
Heimdallr-26B
Melody1437-12B
Melody1437-26B
Melody1437-31B
Nimbz_Gemma-4-Dark-Gemistry-31B
Omega-Evolution-26B
Omega-Evolution-31B
Omega-Evolution-9B
Schattenblume-31B
Serenity-12B
Serenity-26B
Serenity-31B
Swish-26B
TheDrummer_Artemis-31B
TheDrummer_Skyfall-31B
Wicked-Tutu-26B
Wicked-Tutu-31B
Something like that, yeah. It's going to be a blast.
I may just decide to move on tbh. I want to work on newer tunes, not revisit such old ones.
I'd heavily used Heimdallr myself, though since KoboldCPP didn't have a template for G4 (originally, probably has it now) i supplied the one provided and didn't really have any issues, though they said G4 was finicky if the template was wrong.
If it's just a configuration issue, i wonder if extracting to individual layers, replacing the 1 layer that needs replacing and rebuilding would handle it. Though i only deal with GGUF files so not sure how much that helps unless those are adjusted.
Could consider an xdeltas; But like ROMs elsewhere, people don't want/like to do patching.
Regardless, if you feel there's other tunes to do then probably go for it.
Yeah having calculated the costs involved I'm just.............. probably moving on.....
I will be doing different combinations of the prior datasets though. Swish is dark scarlett + wicked tutu. Both 26BA4B and 31B training now.
Well what's actually WRONG with them? I mean, Froopert is my current go-to LLM.
Well what's actually WRONG with them? I mean, Froopert is my current go-to LLM.
Froopert isn't my tune lol
Well what's actually WRONG with them? I mean, Froopert is my current go-to LLM.
Froopert isn't my tune lol
my mistake on that, i scraped on what was 'activity' from FrenzyBiscuit to determine affected models, but it basically says if FB interacted with a model, not if they owned/created it, then filtered the most likely sizes (9B/12B/26B/31B) for Gemma4.
So take my list with a grain of salt :P
It's all good. Shit happens, and it's just one of those things. We appreicate the work and coming forward with the statement that they're less reliable.