SupersonicLabs/Julia-1.5-core

Julia-1.5 Core

Frozen generalist decision model for choice, noul, and score. This repository is the 62,500-step checkpoint selected before the final test. Stage 2 is not part of this release.

Prefer the full package: SupersonicLabs/Julia-1.5 (Core + experts + julia15 runtime). This repo is the Core weights only, same SHA.

The weights are a JuliaDecisionModel on mmBERT-small. They are not a Transformers AutoModel. config.json is the Hub download-stats manifest. The runtime reads julia_config.json, encoder/config.json, and tokenizer/.

Independent matched eval

Same CUDA protocol as Julia-1 (1024/512). On SLCMB editorial core (200×100): Julia-1 39.1% → Julia-1.5 Core 45.0% (+5.9 pp), ECE 0.39 → 0.07. Typed Decisions / Emotion / AG News still favor Julia-1 — see the Julia-1.5 card.

Validation

Measured on the training validation split at the frozen checkpoint (42,120 examples). This is not a separate held-out test report.

Slice	Examples	Accuracy	NLL
Overall	42,120	64.16%	0.791
choice	14,890	62.38%	0.833
noul	15,048	70.43%	0.526
score	12,182	58.59%	1.065

Overall ECE on this split is 0.039. Per-domain numbers are in metrics/validation.json.

Steps	62,500
SHA256	9b6d30c6498324487080b51f84281c571f9d2f2906fd0e219f4f32c80bba70ae
Encoding used in training	1,024 tokens, 512-token head

Domain experts trained on this checkpoint are in Julia-1.5-experts.