--- base_model: unsloth/gemma-4-12b-it license: gemma language: [en] library_name: gguf tags: [gguf, llama.cpp, lm-studio, gemma, gemma-4, qlora, unsloth, methodology, coding-assistant, agent] pipeline_tag: text-generation quantized_by: CodeMonkey1 --- # 🦾 Gemma-4-12B β€” Superpowers Edition ### *A local coding companion that thinks before it types.* Most models hear "build me a tool" and immediately vomit code. **Not this one.** Gemma-4-12B-Superpowers has six engineering disciplines fine-tuned **into its weights** β€” so it reaches for the right method on its own, **with no system prompt, no jailbreak, no babysitting.** --- ## ⚑ What she does differently β€” automatically | You say… | She does… | |---|---| | "Build me an X" | 🧠 **Brainstorms** β€” asks the right questions, weighs 2–3 approaches *before* a line of code | | "It's broken / wrong output" | πŸ”¬ **Root-causes** β€” reproduces & isolates before patching (no guess-fixing) | | "Implement this feature" | βœ… **Test-first** β€” failing test β†’ watch it fail β†’ minimal code | | "Here's a multi-step task" | πŸ—ΊοΈ **Plans** β€” an ordered roadmap before touching files | | "Is it done? Ship it." | πŸ”Ž **Verifies** β€” runs the check, shows the evidence (no "should work") | | "Write a runbook for X" | πŸ““ **Documents** β€” clean SOPs with triggers + red flags | …and she **won't** over-think a one-liner β€” ask "what's 2+2" and you get **`4`**, not a discovery meeting. ## πŸ› οΈ Built for real work β€” WordPress, head to toe ~40% of her training lives in the trenches you actually work in: **PHP** (ACF, hooks, `WP_Query`, WP-CLI), **JavaScript** (Gutenberg blocks, enqueued scripts, jQuery/vanilla), and **CSS** (responsive, CLS-safe fonts, WCAG 2.2). She debugs *your* stack, not toy code. ## πŸš€ Quick start (LM Studio) 1. Download **`gemma-4-12b-it.Q4_K_M.gguf`** β†’ load in LM Studio. 2. **Leave the system prompt EMPTY.** The discipline is baked in β€” that's the whole point. 3. Talk to her like a sharp junior dev. Watch her ask the right questions. *Want her eyes too?* Also grab **`…BF16-mmproj.gguf`** β€” Gemma 4's vision projector. Drop it in the same folder as the text GGUF and LM Studio pairs it automatically; image input works (confirmed describing real photos). Vision weights are stock Gemma 4 β€” untouched by the text fine-tune. ## πŸ”§ Under the hood QLoRA on **`unsloth/gemma-4-12b-it`** via [Unsloth](https://github.com/unslothai/unsloth) β€” r=16, ~1,134 examples (389 discipline + 145 "answer-it-straight" + ~600 general to keep her smart), 2 epochs, **no system message in training** (so the behavior is unconditioned). Methodology adapted from the open-source [superpowers](https://github.com/obra/superpowers) skills (MIT). ## πŸ” v1.1 β€” serving fixes (2026-06-15) - **Chat template corrected** to the canonical Gemma-4-12B-it template. The earlier build shipped a stale template that made LM Studio's engine-protocol runtime throw `Failed to parse input` (HTTP 400) mid-generation. Fixed β€” code-gen eval now runs clean (6/6, no 400s) on q4 and q6. - **Vision restored.** The earlier `mmproj` was mis-converted and crashed on load; replaced with the stock (weight-compatible) Gemma-4-12B projector. Image input confirmed working. ## ⚠️ Status **v1 β€” experimental.** Code-precision A/B eval (q4 vs q6) shows both quants produce correct, idiomatic code across the probe set; the disciplines fire with no system prompt. It's a 12B running locally, so it follows the disciplines well but not frontier-perfectly. Inherits the [Gemma Terms of Use](https://ai.google.dev/gemma/terms). --- *Fine-tuned with stubbornness and ~$0.21 of GPU time. πŸ§ͺ*