Open research evidence · x86-64 Assembly · reproducible

NIYAH Assembly
Evidence Lab

A genuinely trainable 9,616-parameter next-byte neural model implemented in GNU x86-64 Assembly. This site shows a preregistered, independently checkable evaluation—not a hosted LLM or chatbot.

نيّة: تجربة تعليمية قابلة للتحقق بلغة الأسمبلي، وليست نموذج لغة كبيرًا.

Trainable neural weights9,61632-bit float values, native training

Causal context8 bytesNot an instruction-following LLM

Preregistered benchmark24 stories16,805 scored targets

What actually passed

Before downloading the new TinyStories test rows, the evaluation code and row IDs were frozen in commit f7540b9. The first run evaluated official validation rows 1200–1223. The count model was calibrated independently; all scored contexts remain inside their own stories.

Scored targets · NLL ↓	Calibrated count	Native Assembly	Python hybrid
All 16,805	2.258089	2.452823	2.069129
Seen byte IDs (16,799)	2.256635	2.449753	2.067430
Unseen targets (only 6)	6.329330	11.049475	6.827065

Observed, not advertised: The fixed hybrid has 8.37% lower aggregate NLL than calibrated count on this dataset, with a descriptive paired-story bootstrap interval of −0.2123 to −0.1634 nats/byte. But it is worse on the six truly unseen byte targets. This does not establish useful generation, code comprehension, or human impact.

Inspect every test story

Each bar represents the measured count-model NLL minus hybrid NLL per story, calculated from the committed per-story probability loss sums. Positive values mean the hybrid's observed loss was lower.

Loading the machine-produced receipt…

The table above is available even if the interactive receipt cannot load.

Which parts are actually Assembly?

GNU Assembly: native ELF training, forward pass, learned Q/K/V, loss, true-byte probability output, AVX2 dispatch, checkpoint-safe inference and a separate exact-rational calculator.

Python (outside the native model): data retrieval, n-gram baseline, static mixture, testing, numerical oracles and bootstrap statistics. The hybrid is not an all-Assembly generator.

Not provided: instruction-following, semantic understanding, natural-language chat, an LLM superoptimizer, model weights trained on large datasets, cloud inference or a production safety guarantee.

Reproduce on your own Linux/WSL CPU

Uses GNU cc/as/ld, make and Python 3 only for verification. The project source is mirrored in the Files tab and retained at its GitHub commit.

git clone https://github.com/Grar00t/niyah-assembly-model.git
cd niyah-assembly-model
git checkout 1e2e5ed8e5ce6af6106001fe5edfa18e890cc7c8
make verify-all
# Optional: network access only for official TinyStories evaluation
python3 tools/preregistered_eval.py

The source dataset uses TinyStories (CDLA-Sharing-1.0). No downloaded raw story text is redistributed here. Machine-produced numerical statistics, hashes and source tests are available.

Evidence and rights boundaries

The source project has no declared reuse licence yet. Publicly viewable source does not automatically grant permission to reuse it. An owner-chosen licence is needed for downstream redistribution rights. This Space mirrors author-authorized source and reports evidence; it does not apply a new licence.

Human benefit remains NOT MEASURED. A next step is to have a new learner reproduce the benchmark and report task completion and errors; numerical loss is not a substitute for that observation. See the full limitations and first-run evidence.