abuz8studios/arabic-science-packet-a10

Arabic Science Packet A10 (verifiable)

Arabic science word problems (physics, chemistry, biology, earth science) where every answer is a number that can be machine-checked. Built by ABUZ8 LLC for the SUPERBRAIN-1 Arabic brain (https://huggingface.co/abuz8studios/SUPERBRAIN-1).

• 6,021 train / 1,026 test, 27 deterministic template generators, seed 10.
• No LLM-generated content: every question, step-by-step worked solution and gold answer comes from Arabic templates and Python arithmetic in tools/build_packet_a10.py. Chemistry uses a fixed molar-mass table.
• Graded numerically: the last number in the model output is compared to answer_value with relative tolerance tol_rel (0.01). Arabic-Indic digits, Arabic decimal/thousands separators and x 10^n notation are handled.

Subject	Topics	Train	Test
physics	kinematics, newton, energy, electricity, density, pressure, waves	3,345	570
chemistry	molar_mass, moles, molarity (incl. dilution), ideal_gas, stoichiometry	1,338	228
biology	genetics (Punnett mono/dihybrid), growth (doubling)	669	114
earth	atmosphere (lapse rate), seismology	446	76
general	unit_conversion	223	38

Split and leakage

Each generator has 3 phrasings: phrasings 0/1 go to train, phrasing 2 is used only in test, and test parameter tuples are disjoint from all train parameter tuples of the same generator.

From tools/packet_a10_build.log:

leakage train_vs_exam {'exact': 0, 'norm': 0, 'qnorm': 0, 'qhash': 0, 'param_combos': 0}
leakage vs packet_b_math(A9_exam) (n=1190): train {'exact': 0, 'norm': 0, 'qnorm': 0, 'qhash': 0} | exam {'exact': 0, 'norm': 0, 'qnorm': 0, 'qhash': 0}
leakage vs packet_b (n=1343): train {'exact': 0, 'norm': 0, 'qnorm': 0, 'qhash': 0} | exam {'exact': 0, 'norm': 0, 'qnorm': 0, 'qhash': 0}
LEAKAGE=0
packet_a10_train sha256[:16]=5eebbb36f47c6716 | packet_a10_exam sha256[:16]=b18556878e2a4f9d

(data/train.jsonl = packet_a10_train.jsonl, data/test.jsonl = packet_a10_exam.jsonl, byte-identical.)

Verification (tools/check_a10.py, exit 0)

Gold completions must score 100%, shuffled (other-item) answers must score < 5%, plus parser unit cases.

[train] n=6021 gold_completion=6021/6021 (100.00%) gold_answer=6021/6021 (100.00%) gold_arabic_digits=6021/6021 | shuffled_completion=10/6021 (0.17%) shuffled_answer=10/6021 (0.17%)
[exam] n=1026 gold_completion=1026/1026 (100.00%) gold_answer=1026/1026 (100.00%) gold_arabic_digits=1026/1026 | shuffled_completion=2/1026 (0.19%) shuffled_answer=2/1026 (0.19%)
[unit] parser cases passed 5/5
SELF-TEST PASS

Fields

task, src, subject, topic, template, prompt, completion, answer, unit, answer_value, tol_rel, qhash. prompt ends with الحل:; completion is the worked steps followed by الجواب: <number> <unit>. answer is the rounded number string, answer_value the exact float.

Tools

• tools/build_packet_a10.py - the generator (writes to $A10_DIR, default: its own folder; the external leakage check needs packet_b_math.jsonl / packet_b.jsonl beside it and reports LEAKAGE=-1 "cannot certify" if they are absent).
• tools/check_a10.py - scorer score(pred_text, item) + self-test (python tools/check_a10.py).
• tools/a10_eval.py - runs the test split against a llama.cpp llama-server (--url, --name) and reports accuracy per subject/topic.
• tools/packet_a10_build.log - build log.

Related

• Model: abuz8studios/SUPERBRAIN-1
• abuz8studios/arabic-code-packet-a11 - Arabic-instruction Python tasks (verifiable)
• abuz8studios/arabic-sarf-packet-b - Arabic morphology (sarf)

Model scores

None yet. Baselines for SUPERBRAIN-1 will be added here once measured.