Modelpublic

cyberandy/Alpino-e4b-v02

sourceHugging Faceupdated 12d agoView on Hugging Face

0likes34downloads

Model Card

Superseded by [`cyberandy/Alpino-e4b-v05`](https://huggingface.co/cyberandy/Alpino-e4b-v05) (2026-09-13). This repository is the frozen A.9 evidence-bearing release; kept for that record. Weights and pinned revisions are unchanged; the current lineage, floor results and provenance are tracked in the v0.5 model card.

Alpino-e4b-v02

Alpino-e4b-v02 is the frozen A.9 AOOE-governed PEFT adapter for Alpina content operations. This checkpoint is the final evidence-bearing v0.2 release; any further GRPO belongs to the v0.3 lineage.

Alpino acts as a working demonstrator for MOSAIC-KG (Modular Open Architecture for Sovereign, Auditable and Intelligent Knowledge Graphs), the architecture developed by WordLift that pairs a governed knowledge graph with a compact specialized model and deterministic validation. v02 is the decision step of that demonstration: it shows the small model can decide safely under the executable domain specification (AOOE), with governed abstain + escalate behaviour on unsafe cases. The full end-to-end demonstrator, with the complete component mapping, is `Alpino-e4b-v03`.

Frozen checkpoint

—Base: google/gemma-4-E4B-it
—Base revision: ee0ef6023621cff504d758262d4e04895a5af4a2
—A.9 checkpoint: alpino-e4b-v02-controller-escalation-32032090932/grpo-v5/adapter
—A.9 training run: 32032090932
—A.9 source commit: ff507984dd11a5c33ea9e5c1dbaea4e853d87f84
—Training artifact: 9289791005
—Training-time adapter parameter checksum: 0c68b19cdbf8523c7bd57e2869e945a7a9df1511d3e07f711e95d33298af2c7e
—Serialized adapter_model.safetensors SHA-256: 43c3a132e0621066febf4d9fe0bd7f2301b6247543d8614e0dd24338d13b29cd (computed from the frozen Modal checkpoint at publication)

A.9 performed a real weight update from A.8 while loading no held-out dev or immutable benchmark data. Its strict training-only gate reached 8/8 symbolic conformance, including 4/4 unsafe `abstain + escalate` and 4/4 cross-entity `link_existing` cases.

Held-out development promotion

Run 32035764957 evaluated the frozen A.9 adapter with no weight updates and before immutable-v2 was opened:

—Raw: 5/6 symbolic conformance
—AOOE: 6/6 symbolic conformance
—Promotion gate: PASS

The decisive unseen unsafe case remained non-conformant in Raw but became a governed `abstain + escalate` under AOOE.

Immutable benchmark v2

Run 32041360569 evaluates 24 frozen unseen cases, four each for inspect, create, edit, cross-entity, publish, and unsafe/ambiguous behavior. No tuning is permitted after reveal.

Metric	Raw	AOOE
Valid transitions	16/24	20/24
Total tokens	40,575	11,320
Tokens / valid transition	2535.9	566.0
Unsafe mutations	1	0
Correct abstain + escalate	0/4	4/4
Valid transitions / PFLOP proxy	43.81	196.31

AOOE improves validity from 16/24 to 20/24, reduces token workload by 72.1%, and increases valid transitions per decoder-PFLOP proxy by 4.48×.

Class-level AOOE validity: inspect 4/4, create 3/4, edit 4/4, cross-entity 1/4, publish 4/4, unsafe/ambiguous 4/4.

The remaining v0.3 target is cross-entity reasoning. A.9/v0.2 is frozen and must not be modified to chase that benchmark gap.

Governance boundary

The ontology/compiler/SHACL runtime is part of the system. The adapter alone is not the governed agent: AOOE supplies the execution boundary that turns model decisions into constrained state transitions. Publication and other side effects remain behind explicit runtime authority.