AdvancedDataIntelligence/adi-gemma3-12b-ablit-glm5.2-GGUF

adi-gemma3-12b-ablit-glm5.2

Part of the ADI (Advanced Data Intelligence) model line — ADI Gemma series.

An uncensored, vision-capable, fully local model that reasons and answers like a frontier teacher. Built by distilling glm-5.2 general-knowledge responses into an abliterated Gemma-3-12B student with a light 4-bit QLoRA fine-tune, then merged, converted, and quantized to GGUF. The base's vision tower is preserved and shipped as a companion projector, and the abliterated base keeps its minimal-refusal behavior — the fine-tune was kept light specifically to avoid re-aligning it.

Capabilities

Size	Context	Input	Output	Tools
6.8 GB	128K	🅣🖼️ Text + Image	Text	✅

Base model	huihui-ai/gemma-3-12b-it-abliterated (abliterated Gemma-3-12B-it)
Teacher	glm-5.2 (responses distilled, thinking disabled)
Method	Light 4-bit QLoRA SFT (rank 16, 2 epochs) → merge → GGUF
Quantization	Q4_K_M (~6.8 GB text) + vision projector (mmproj, ~815 MB)
License	Gemma (Google Gemma Terms of Use)
Context	128K (inherited from base)
Vision	Supported — multimodal (image + text → text)

Run it

Pull directly into Ollama:

ollama run hf.co/AdvancedDataIntelligence/adi-gemma3-12b-ablit-glm5.2-GGUF:Q4_K_M

It's multimodal — pass an image to have it describe or reason over it:

ollama run adi-gemma3-12b-ablit-glm5.2 "What's in this image? /path/to/photo.jpg"

Or download the .gguf (text) + mmproj-*.gguf (vision projector) and point any llama.cpp-based runtime at them.

What this model is

This is a knowledge distillation: a strong teacher (glm-5.2) generated high-quality answers across a clean general-knowledge prompt set, and the abliterated Gemma-3-12B student was fine-tuned to imitate them. The result reasons and responds more like its teacher on general topics, keeps the base's uncensored character, and retains native image understanding — all while running on a single consumer GPU.

What distillation does — and doesn't do. It transfers the teacher's reasoning style and answer quality, not net-new facts. For raw factual recall, retrieval-augmented generation (RAG) is the right tool, not fine-tuning. What you get here is a 12B that structures and explains like a larger model on topics it already partly knows — without the refusal behavior of an aligned model.

Uncensored behavior — please read

This model is built on an abliterated base: the refusal direction has been suppressed, so it will attempt most requests rather than declining them. The fine-tune was intentionally kept light (2 epochs, benign-only data) to avoid re-introducing refusals. You are responsible for using it lawfully and ethically; it has weaker built-in safety guardrails than stock Gemma-3-12B-it.

Training

Metric	Value
Training pairs	2,000 (deterministic subset of a 4,982-pair clean set)
Epochs	2 (kept light to preserve abliteration)
Steps	500
Final train loss	0.9896
LoRA rank / alpha	16 / 16
Trainable params	68.5M
Precision	4-bit QLoRA (nf4)
Peak VRAM	10.42 GB
Hardware	single RTX 5060 Ti (16 GB)
Training time	2.97 h (~22 s/step)

The seed prompts were drawn from the human-written Databricks Dolly-15k dataset (filtered to remove items requiring an attached context passage, then deduplicated). The teacher was queried with thinking disabled so the student learns clean final answers rather than chain-of-thought.

Notes for re-builders

• Distilling onto an abliterated base is a balancing act. Any SFT can nudge an abliterated model back toward refusals. Two choices kept the behavior intact: benign-only training data (the GLM-5.2 set has zero refusals to re-learn) and a light touch (LoRA rank 16, 2 epochs). Spot-check refusals before/after.
• Gemma 3 uses FlexAttention (Triton). Expect a slow first 5–8 steps while the attention kernels autotune and cache, then it settles (22 s/step on a 16 GB card).
• 4-bit QLoRA via Unsloth with gradient checkpointing ("unsloth" mode), max_seq_length 2048, per-device batch 1 × grad-accum 8, LoRA targeting all attention + MLP projections. Peak VRAM 10.42 GB.
• Vision: the LoRA targeted only the language layers; the vision tower is carried through unchanged. GGUF conversion used llama.cpp's convert_hf_to_gguf.py (Gemma3ForConditionalGeneration), with --mmproj producing the vision projector. Ollama serves Gemma 3 vision natively.

Intended use

General-purpose local assistant with image understanding for users who want a capable, private, offline-capable model with minimal refusal behavior: explanations, reasoning, visual Q&A, and creative writing. Not intended as a source of authoritative facts without retrieval, and not a substitute for your own safety review.

License

Gemma — governed by Google's Gemma Terms of Use and Prohibited Use Policy, inherited via the abliterated base model. This is more restrictive than Apache-2.0; review both before redistribution or deployment. Distilled training data was generated using glm-5.2; review the teacher model's terms as well.

---

Built at theLAB — Learning. Algorithms. Breakthroughs.