SUPERBRAIN-1 — Arabic-first root reasoning + Arabic math (Fanar-1-9B)

SUPERBRAIN-1 is a 9B Arabic model by ABUZ8, built on QCRI/Fanar-1-9B. It is trained to reason through Arabic morphology (ṣarf): find the root of a word, build any verb Form I–X from a root, conjugate, derive. On top of that, it solves Arabic math word problems step by step.

It is a completion model, not a chat model. Give it one of the prompt formats below and read the first line of the output.

Files

File	Version	What it is
superbrain1-run6-q5_k_m.gguf	v2 (current, 2026-10-12)	roots + Quranic words + Arabic math
superbrain1-q5_k_m.gguf	v1 (2026-10-06)	morphology specialist (best verb forms)

Results

All scores use greedy decoding on the Q5_K_M GGUF served by llama.cpp. Evaluation files are in eval/.

Packet B — held-out test. 1,343 items over 258 roots that were excluded from all training data: real Quranic words (gold roots from the Quranic Arabic Corpus) and modern words from the Arramooz lexicon. The exam is published as a dataset: abuz8studios/arabic-sarf-packet-b.

Arabic math. All 1,190 problems of the Arabic GSM8K test split (Omartificial-Intelligence-Space/Arabic-gsm8k-v2), with exact match on the final number. The test questions were hash-checked so that none of them reached training.

	Fanar-1-9B base	v1 (SFT)	v2 (run6)
Root extraction, bare question	59.5%	66.8%	78.0%
Root extraction, few-shot	43.2%	67.9%	78.6%
Verb Forms I–X (428 items)	16.6%	94.9%	88.8%
Arabic GSM8K (1,190)	—	—	63.6%
85-item gate (v3)	61.2%	96.5%	95.3%

Math rose from 21.9% to 63.6%. The checkpoint before the math stage (Quranic-word round) scored 21.9% on the same 1,190 problems. 300 GRPO steps with an exact-answer reward took it to 63.6%.

v2 root accuracy by root class (bare question):

Class	Accuracy
sound	90.7%
doubled	68.8%
hollow	68.6%
assimilated	64.8%
hamzated	63.5%
defective	48.1%

v1 scored 79/56/60/44/51/45% on the same classes.

Generalization probe. The root خدم was banned from all training data. Asked for the root of مستخدمين ("users"), both v1 and v2 answer خدم under all three prompt formats.

Prompt formats

Root (bare): ما جذر كلمة «مستخدمين»؟\nالجذر: → خدم

Verb form: الجذر: درج — الصيغة العاشرة\nالفعل: → استدرج

Conjugation: الفعل: نجم | الزمن: المضارع المجزوم | الضمير: هما\nالتصريف: → ينجما

Math:

المهمة: حل السؤال التالي خطوة بخطوة ثم اكتب الجواب النهائي رقمًا.
سؤال: ...
الحل:

→ The model writes worked steps, then الجواب: <number>.

Usage (llama.cpp)

llama-server -m superbrain1-run6-q5_k_m.gguf -ngl 99 -c 4096 --port 8099
curl -s http://127.0.0.1:8099/completion -d '{"prompt":"ما جذر كلمة «مستخدمين»؟\nالجذر:","n_predict":16,"temperature":0}'

How v2 was trained

Every stage is kept only if it beats the previous champion on Packet B without losing more than 2 points elsewhere.

1. Supervised fine-tuning (v1) on 600,000 generated gold items from 5,955 roots: root extraction, conjugation, verb forms, form naming, derivation. QLoRA r64, loss on the answer only.
2. GRPO rounds on roots and morphology (champion "rl1f"), with an exact-match reward checked by a program.
3. GRPO on Quranic words (355 hard items, 400 steps).
4. GRPO on Arabic math (662 Arabic GSM8K train problems the model sometimes got right, 300 steps, 8 samples per prompt), with an exact-number reward.

Every reward is a program check, either a root oracle or an exact number. No AI judge is used.

Limitations

• v2 traded some verb-form accuracy for roots and math (94.9% → 88.8%). Form VI is the weak spot at 55%. Use v1 if you only need verb forms.
• About one root in five is still missed. Defective roots (final و/ي) are the hardest class at 48%. Errors also come from clitic ambiguity out of context.
• Single words only. There is no sentence context and no disambiguation.
• Not chat-tuned. Free-form questions get base-model behaviour.
• Quranic spelling. Quranic words are tested in modern spelling, converted from Uthmani script.

Data and licenses

• Base model: QCRI Fanar-1-9B, Apache-2.0.
• Morphology training data was generated with Arramooz, Qutrub and PyArabic (all GPL). Those tools are not distributed here. Only generated question–answer pairs were used for training.
• Math: Arabic GSM8K v2 (Apache-2.0). Only the train split was used for training; the test split was used only for evaluation.
• Evaluation: Quranic Arabic Corpus (GPL).

Thanks to QCRI for Fanar, Taha Zerrouki for Arramooz, Qutrub and PyArabic, Kais Dukes for the Quranic Arabic Corpus, and Omartificial Intelligence Space for Arabic GSM8K.

Experiment log 2026-10-13

Two more training lines were run and graded. None of them replaced v2 (run6). No new weights were uploaded. This section shows the recipes and scores only.

Every run uses the same evaluation as above: Q5_K_M GGUF, greedy decoding, llama.cpp. A run is promoted only if its Packet B bare-root score is at least run6's (0.7803), it loses no more than 2 points on verb forms or math, and the 85-item gate passes.

Run	Body	Packet B root, bare	Packet B root, few-shot	Verb forms (morph)	Arabic GSM8K (math)	85-item gate (v3)	Verdict
run6_math_grpo (v2, champion)	Fanar-1-9B	0.7803	—	0.8879	0.6361	—	kept
g2f_sft	Fanar-1-9B	0.6448	0.6612	0.9369	0.6950	0.9412	not promoted
g2f_rl	Fanar-1-9B	0.7574	0.7661	0.9346	0.7092	0.9765	not promoted
g2f_math	Fanar-1-9B	0.7552	0.7563	0.9393	0.7647	0.9647	not promoted
g4_sft1	Gemma-4-E4B	0.6842	0.6852	0.7874	0.4202	0.9529	not promoted
g4_rl	Gemma-4-E4B	0.7443	0.7650	0.7570	0.4218	0.9412	not promoted
g4_math (best G-line)	Gemma-4-E4B	0.6863	0.6842	0.7757	0.5176	0.9412	not promoted
g4_ta	Gemma-4-E4B	0.7399	0.7716	0.7290	0.5134	0.9176	not promoted

The champion row lists only the three numbers the promotion rule compares against. Every graded run beats the Fanar-1-9B base on Packet B bare roots (0.5945).

Plain summary. No run beat run6 on the bare-root bar of 0.7803. The closest was g2f_rl at 0.7574. The G2 line got the best Arabic math score so far, 0.7647 (g2f_math, against run6's 0.6361), and better verb forms, 0.9393 (against 0.8879). It was still not promoted, because the rule requires the bare-root score first. The Gemma-4-E4B line (G-line) stayed below run6 on roots, verb forms and math.

Recipes (one line per run)

G2 line: Fanar-1-9B base, retrained from scratch on a new curriculum

• g2f_sft: QLoRA SFT from the Fanar-1-9B base on g2_sft_mix (78,202 rows: A7 60k + A9 mix + A8). The 1,292 RL-pool roots and the 269 dev roots were checked to be absent. 1 epoch, r64, lr 1e-4, max length 768, batch 4 × accum 16. Final train loss 0.3747.
• g2f_rl: GRPO from g2f_sft on roots it had never seen (rl_pool_a7; 2,163 prompts kept after filtering), 300 steps, exact-root reward.
• g2f_math: GRPO from g2f_rl on Arabic GSM8K train problems (packet_a9; 1,200 prompts kept out of 3,000 candidates), 300 steps, 8 generations, batch 8 × accum 4, max completion 384, exact-number reward.

G-line: Gemma-4-E4B body (the base model also takes images and audio)

• g4_sft1: SFT from the raw Gemma-4-E4B on sft_a9_mix (16,645 rows), r64, lr 1e-4, max length 768, batch 1 × accum 64.
• g4_rl: GRPO from g4_sft1 on 1,103 roots that were absent from its SFT data (3,000 prompts), 600 steps, batch 16, 8 generations, accum 2, max completion 48.
• g4_math: GRPO from g4_sft1 on packet_a9 math (1,200 prompts), 250 steps, batch 4, 8 generations, accum 8, max completion 384.
• g4_ta: weight arithmetic, with no further training: g4_rl + g4_math − g4_sft1 (both RL deltas at full size). The rule was fixed before scoring, and the model was graded once.

Every reward is a program check, either an exact root or an exact number. No AI judge is used.

What we learned. g4_ta shows the weakness of the Gemma line. Its verb forms collapsed on Forms III (25%), VI (2.5%) and VIII (55%). The roots-GRPO stage traded verb forms for roots. The G2 Fanar line kept verb forms high (Form VI 92.5% in g2f_rl and g2f_math). It reached 75.7% on bare roots, which is still short of run6's 78.0%.