asketeddy/gooo-ir-operator-tiny-v1

Gooo IR operator tiny v1

A 12,728-parameter experimental classifier for eight typed binary IR operations: add, subtract, multiply, less_than, less_equal, equal, and, or. It maps a bounded English/Korean instruction to probabilities. The Go bridge validates operand types and identifiers before constructing a closed IR node and expression. It does not generate arbitrary source or a complete program.

Origin: independently initialized MLP, not Laya weights or a Laya fine-tune. Laya is the research baseline for future encoder/head experiments. All training material is generated public synthetic templates. Code, data and these independently trained weights are MIT licensed.

Evaluation and completeness

2,048 examples: 1,536 training, 256 calibration, 256 test. Entire prompt templates are kept in one split; the test has 32 templates with eight variable configurations each, 128 English/128 Korean examples and 32 examples per operation. The effective independent intent/template coverage is narrower than 256 rows. No test data was used to fit weights, checkpoint selection, temperature or abstention thresholds.

Variant	Correct / test rows	Test accuracy	Packed weights	Resident tensor arrays	Hot Go prediction
FP32	256 / 256	100%	50,912 B	50,912 B	7.86 us
PTQ ternary	245 / 256	95.70%	2,759 B	12,896 B	9.34 us
QAT ternary	242 / 256	94.53%	2,759 B	12,896 B	9.30 us

Go matched all 96 Python parity rows within absolute tolerance 1e-4. The frozen Go audit contains label/language/template breakdowns, calibration and selective acceptance, model/input/source hashes, typed IR checks and latency. PTQ accepted 251/256 and was correct on 242 accepted rows. QAT accepted all 256 despite 14 errors; confidence is not a guarantee of semantic correctness. Boolean conjunction/disjunction in held-out English templates caused most ternary errors. Unseen operation families, free-form prompts, ambiguity, malicious inputs and real developer workloads have not been evaluated.

All hot prediction variants used 0 B/op and 0 allocations/op in this Apple M4/Go 1.27 measurement. Ternary was approximately 19% slower than FP32. Tensor totals exclude metadata, object headers, loader temporaries and, for ternary, eight additional scale bytes. One reusable worker workspace is 1,248 B; the prediction struct is 80 B. Cold CLI median peak RSS was about 5.2–5.4 MiB over three processes per variant. FP32's first startup took 385 ms; its median was 8.33 ms. PTQ/QAT medians were 3.35/3.24 ms. These small samples are not a deployment latency guarantee.

Quantization

The ternary alphabet has theoretical log2(3) = 1.585 bits. Packing five trits per byte uses 1.6 physical bits per matrix weight, plus biases, scales and metadata. The Go runtime decodes once to contiguous int8 matrices. Resident RAM is not 1.58-bit packed storage. QAT used floating-point master weights with straight-through ternary training; this is a small MLP, not a BitNet Transformer or an 8-bit activation implementation. PTQ is derived from FP32; QAT is a separate training arm. This pilot did not show QAT outperforming PTQ.

Use with Go

Get the runtime from the public GitHub repository. Download a variant's model.json and sibling weights.bin from runs/pilot-mps-20260930-v1/models/. Then run:

go run ./cmd/gooo-decision --model runs/pilot-mps-20260930-v1/models/fp32/model.json <<'JSON'
{"schema":"gooo/tiny-ir-decision-request/v1","text":"Operands: a, b. Combine both values by summation.","left":{"name":"a","type":"Int"},"right":{"name":"b","type":"Int"}}
JSON

The runtime rejects invalid UTF-8, empty instructions and instructions above 512 UTF-8 bytes. It never silently truncates input. Model loading rejects digest, shape and encoding mismatches. A type mismatch or calibrated low confidence produces an abstention. Deterministic ordinary Gooo compilation remains the fallback path; this experimental bridge is not integrated into production Gooo.

Training resources and provenance

Training used PyTorch 2.14 on Apple M4 MPS, ten GPU cores, 16 GiB unified RAM, 120 epochs per FP32/QAT arm and batch size 256. Measured loop time was 2.62 s for FP32 and 1.02 s for QAT, including calibration and CPU scoring. The first arm includes warm-up/cache costs; this is not evidence QAT training is faster. Process-lifetime peak RSS was approximately 422–424 MiB. End-of-epoch sampled MPS allocation/driver maxima were 2.06/53.18 MB. GPU utilization percentage, continuous GPU memory peak and kernel-only timing were not measured.

This Hub bundle contains an explicit 17-file allowlist: synthetic data and manifest, model contract/generator, frozen training source/preexecution, filtered summary, machine card, Python parity, external Go audit, and three model pairs. Tokens, private repository material, local filesystem paths and hardware identifiers are excluded. SHA checks and pattern scans are integrity/hygiene checks, not proof of privacy or quality outside this finite experiment.

Research sources: Laya, BitNet b1.58 paper, PyTorch MPS.