SUPERBRAIN-1-G4: Arabic roots and morphology on Gemma-4-E4B

The best run of SUPERBRAIN-1's G-line (run name g4_morph), built on google/gemma-4-E4B (Apache-2.0). It's a smaller, multimodal-capable sibling of the Fanar-based SUPERBRAIN-1. It is a completion model, not a chat model; the prompt formats are in the SUPERBRAIN-1 card.

Files

File	What it is
superbrain1-g4-morph-q5_k_m.gguf	g4_morph, Q5_K_M
mmproj-superbrain1-f16.gguf	Image projector (mmproj) from the sibling run g4_sft1 (same Gemma-4-E4B body). g4_morph has none of its own, and this pairing hasn't been graded on images.
lora/	g4_morph's LoRA adapter (r=32, alpha=64). Its config names google/gemma-4-E4B as the base, but it was actually trained on top of g4_ta (see Lineage), so it reproduces g4_morph only on top of g4_ta.
verdict.json	Full eval output

Scores (verdict.json, Q5_K_M, greedy, llama.cpp)

Metric	g4_morph	Gemma-4-E4B base
Gate v3 (85 items, bar 0.8471)	0.9647	0.5647
Packet B root, few-shot	0.7891	0.5617
Packet B root, bare	0.7694	0.3945
Verb forms	0.8598	0.0023

Verb forms by form: I 0.8485 · II 0.9796 · III 0.8182 · IV 0.9437 · V 0.9545 · VI 0.325 · VII 0.8571 · VIII 0.7586 · X 1.0

Bare roots by root class: sound 0.8841 · doubled 0.7662 · hollow 0.7581 · assimilated 0.6338 · defective 0.4805 · hamzated 0.4595

Honest notes: Form VI is weak at 0.325, and so are hamzated and defective roots. The Fanar line is better on bare roots (run4_a8_quran: 0.7825) and verb forms (soup_g2f_run6: 0.9276). G4's edges are few-shot roots, the highest of any run, and a smaller multimodal body.

Lineage

gemma-4-E4B → g4_sft1 (SFT on sft_a9_mix) → g4_rl (GRPO roots) and g4_math (GRPO math) → g4_ta (= g4_rl + g4_math − g4_sft1, weight arithmetic, no training) → g4_morph

Recipe

• Base: g4_ta (merged 16-bit)
• Data: build_grpo2_data.py --verdict g4_ta/verdict.json --data packet_a7.jsonl --candidates 8000 --keep 3000 → grpo_g4_morph.jsonl (aimed at the weakest verb forms and root classes)
• Method: GRPO with an exact-answer program reward (no AI judge): train_grpo_vl.py --max-steps 600 --bs 8 --gens 8 --accum 4 --max-comp 64 --gemma4-cache
• LoRA: r=32, alpha=64
• Eval: gate v3 + Packet B (258 roots held out of all training) on Q5_K_M

About the lab

ABUZ8 is a one-person lab. Ahmad builds it alone, with no team and no funding, and has spent the last year living inside AI work. The long-term goal is the SUPERBRAIN fusion: Arabic-root reasoning, tool use and memory (Qadir), fast local inference (FAIE) and new architecture experiments, all joined in one open system.

• GitHub: https://github.com/abuz8studios-ship-it
• Site: https://abuz8ai.com
• X: https://x.com/AbuZ8Studios