SUPERBRAIN-1-lab: every graded run

This repo holds every training run behind SUPERBRAIN-1. Each run sits under runs/<run>/ with its Q5_K_M GGUF, LoRA adapter (where one exists) and verdict.json. Nothing here is promoted. The champion and the promotion rule are documented in the main repo.

Scores (from each run's verdict.json, Q5_K_M, greedy, llama.cpp)

Run	Body	Root bare	Root few-shot	Verb forms	Gate v3
soup_g2f_run6	Fanar	0.7814	0.7880	0.9276	0.9529
run6_math_grpo	Fanar	0.7803	0.7858	0.8879	0.9529
run4_a8_quran	Fanar	0.7825	0.7847	0.8902	0.9529
g2f_rl	Fanar	0.7574	0.7661	0.9346	0.9765
g2f_math	Fanar	0.7552	0.7563	0.9393	0.9647
run7b_a10_gentle	Fanar	0.7464	0.7574	0.8668	0.9412
soup50	Fanar	0.7311	0.7399	0.9463	0.9765
g2f_sft	Fanar	0.6448	0.6612	0.9369	0.9412
g4_morph	Gemma-4-E4B	0.7694	0.7891	0.8598	0.9647
g4_rl	Gemma-4-E4B	0.7443	0.7650	0.7570	0.9412
g4_ta	Gemma-4-E4B	0.7399	0.7716	0.7290	0.9176
g4_sft1	Gemma-4-E4B	0.6842	0.6852	0.7874	0.9529
g4_math	Gemma-4-E4B	0.6863	0.6842	0.7757	0.9412

Fanar-1-9B base, for reference: root bare 0.5945.

Lineage

• Main Fanar chain: Fanar-1-9B → E3 SFT (checkpoint-750) → rl1f (GRPO on unseen roots) → run4_a8_quran → run6_math_grpo
• G2 retrain: Fanar-1-9B → g2f_sft → g2f_rl → g2f_math
• Soup: soup_g2f_run6 = 0.5 × g2f_math + 0.5 × run6_math_grpo (soup_meta.json, "rule: fixed before scoring")
• G-line: gemma-4-E4B → g4_sft1 → g4_rl and g4_math → g4_ta (= g4_rl + g4_math − g4_sft1, no training) → g4_morph

Every other run (soup50, rl1f, rl3f, run7_a10, run7b_a10_gentle, g2f_sft, g2f_rl, g2f_math, g4_sft1, g4_rl, g4_math, g4_ta) is an intermediate stage or a weaker variant of these, kept for transparency.

All runs share one eval: gate v3 (85 items, bar 0.8471) and Packet B (258 held-out roots: root bare and few-shot, plus verb forms), on Q5_K_M with greedy decoding in llama.cpp. Every reward is a program check; no AI judge is used.

Recipe: run4_a8_quran (best bare roots)

• Base: rl1f
• Data: build_packet_a8.py builds Packet A8 from the Quranic Arabic Corpus: words with one gold root (homographs dropped) and bare 3MS perfect verbs with their gold form. Packet B roots and words and the gate roots are banned and checked to zero. Seed 1408. The GRPO set is grpo_run4_a8_quran.jsonl.
• Method: GRPO (Unsloth/TRL, loss "dapo", beta 0.0), exact-root reward
• Hyperparams: LoRA r=32, alpha=64; 400 steps; batch 16
• Scores: root bare 0.7825, few-shot 0.7847, verb forms 0.8902, gate 0.9529
• Strengths: the highest bare-root accuracy of any run
• Weaknesses: no math training

Recipe: run6_math_grpo (math)

• Base: run4_a8_quran
• Data: build_packet_a9.py → packet_a9.jsonl (Arabic GSM8K train split), GRPO set grpo_run6_math.jsonl
• Method: GRPO, exact-number reward
• Hyperparams: LoRA r=32, alpha=64; 300 steps; batch 8
• Scores: Arabic GSM8K 63.6% (in the main repo's eval/math_arabic_gsm8k_run6.json); root bare 0.7803, few-shot 0.7858, verb forms 0.8879, gate 0.9529
• Weaknesses: verb Form VI 0.55

Recipe: soup_g2f_run6 (balanced)

• Method: linear weight average, alpha 0.5, of the merged 16-bit g2f_math and run6_math_grpo. No training.
• Scores: root bare 0.7814, few-shot 0.7880, verb forms 0.9276 (Form VI 0.90), gate 0.9529
• Weaknesses: just behind run4 on bare roots. Its verdict file has no math score.

Recipe: g4_morph (best G-line, Gemma-4-E4B)

• Base: g4_ta
• 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: train_grpo_vl.py --max-steps 600 --bs 8 --gens 8 --accum 4 --max-comp 64 --gemma4-cache
• Hyperparams: LoRA r=32, alpha=64
• Scores: gate 0.9647, root few-shot 0.7891 (the top few-shot of any run), bare 0.7694, verb forms 0.8598
• Weaknesses: Form VI 0.325, Form VIII 0.7586. Published separately as SUPERBRAIN-1-G4.

Not published

• Qadir qadir_v1_tools_f: same LoRA setup as qadir_v1_tools (r=64, alpha=128), but its changes aren't documented and it has no eval. Unpublished until it's evaluated.

Licenses

• QCRI/Fanar-1-9B: Apache-2.0. Applies to all Fanar-line runs.
• google/gemma-4-E4B (g4_* runs): Apache-2.0 (license). The best G-line run is in SUPERBRAIN-1-G4.

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 (SUPERBRAIN), tool use and memory (Qadir), fast local inference (FAIE) and new architecture experiments (MUTHALLATH), all joined in one open system.

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