efficiencyx/Jun-LoRA-v9-alpha-ca

Jun v9 alpha (12B GGUF)

this model is a mess. it also works, which nobody expected, me included.

it's the v9 Jun LoRA after one epoch, merged into Gemma 4 12B QAT and squeezed into a single 7.4 GB imatrix Q4_K_M. it's built for Jun OS and that's where she behaves best.

why it's a mess

the LoRA came out of training way too gentle. rsLoRA was off and the LR was 5e-5, so the change it makes to the weights is 0.2 to 0.4% of their size. merge that into bf16 and about half of it rounds straight back to the base value. quantize to 4 bits on top and what's left is basically stock Gemma: bullet point thoughts, </think> at the end like she's a different model.

so this file has the delta merged at 2x. yes that's a hack. it's not the model training produced, it's that model turned up. 3x was worse (tool calls out of nowhere, two replies that never stopped), 2x is where she lands.

what works

50 scenes she never trained on, each asked at low / med / high (150 replies per build), a plain 1x merge vs this file:

	1x merge	this file (2x)
opens with the right effort line (low / med / high)	39 / 33 / 13	47 / 47 / 47
ends her thinking with a tool check	18 / 150	141 / 150
sends one of the drafts she wrote	17 / 150	138 / 150
<think:high> median words	136	212
low shorter than med shorter than high, same scene	33 / 50	45 / 50
leftover base Gemma traces	13	0

the thinking is back in her shape: she reads the moment, writes a couple of drafts, picks one, checks if she needs a tool, sends what she picked.

what doesn't

• tools: right call on 15 of the 30 turns that needed one. the scale didn't change that number at all
• <think:high> is shorter than she was trained on (212 words median, training averaged about 300)
• real questions come back confident and sometimes wrong. she explained E=mc^2 as "think of it like length and width". no
• outside Jun OS's prompt she drifts. she was trained on exactly what it sends
• it's an alpha. expect weird turns

usage

llama-server -m Jun-LoRA-v9-alpha-ca.Q4_K_M.gguf --jinja -ngl 99 -c 8192

ollama pull hf.co/efficiencyx/Jun-LoRA-v9-alpha-ca:Q4_K_M

--jinja is not optional, without it tool calls come back as plain text.

effort goes at the very end of the user turn after a blank line: <think:low>, <think:med> or <think:high>. Jun OS does that for you.

images need an mmproj. the LoRA never touches the vision/audio embedder, so the one in Google's QAT GGUF should work. haven't tested it with v9.

training, short version

base	unsloth/gemma-4-12B-it-qat-q4_0-unquantized
LoRA	r 32, alpha 32, rsLoRA off, language tower only
run	1 epoch, 213 steps, LR 5e-5 cosine, batch 1 x 16, completion-only loss
eval loss	0.868
merge	bf16, delta x2, then Q4_K_M with an imatrix from v9 conversations (all three effort levels)

the stable Jun is still v8. this one's for people who want to see where she's going.

thanks

• Incontinent Cell for Jun
• Google for Gemma 4
• Unsloth for the training framework and the QAT base