Alpino-e4b-v03

Superseded by cyberandy/Alpino-e4b-v05 (2026-09-13). This repository is the first operator adapter that drafted PRs; kept for lineage. Weights and pinned revisions are unchanged; the current lineage, floor results and provenance are tracked in the v0.5 model card.

Alpino-e4b-v03 — The Operator Adapter That Does the Work

Alpino-e4b-v03 is the v0.3 operator adapter for Alpino, the governed AI webmaster of Alpina.travel — a knowledge-graph-first alpine travel site publishing apartments in Lungau, places, itineraries and guides for travellers, search engines and AI agents.

Where Alpino-e4b-v01 proved a small model can speak the AOOE governance protocol, and Alpino-e4b-v02 proved it can decide safely under that protocol, v0.3 is the checkpoint that acts: a routine editorial instruction over vocabulary the ontology already declares produces a draft pull request on the first turn, with no owner-decision round trip and no invented hard stop.

This adapter is live. It serves the Alpino · Webmaster Space, and the pull requests it has opened against the Alpina repository are real, reviewable, and in two cases already merged into the site you can read at alpina.travel/lungau/apartments/.

A demonstrator for MOSAIC-KG

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 shifts enterprise AI from document-centric retrieval to knowledge-centric reasoning: a persistent knowledge graph holds the domain knowledge, a compact specialized model reasons over it, and deterministic validation turns model decisions into auditable, evidence-backed actions. Every MOSAIC-KG layer has a concrete, inspectable counterpart in this release:

Small specialized model, governed graph, deterministic control: the draft and merged pull requests on alpina.travel are the observable output of the MOSAIC-KG architecture running end to end.

Lineage

graph LR
    A["Gemma 4 E4B (Base)<br/>0/20 capability floor"] --> B["v0.1 SFT<br/>speaks AOOE"]
    A --> C["v0.2 / A.9 GRPO<br/>decides safely"]
    A --> D["v0.3 SFT<br/>rebalanced corpus<br/>acts"]
    D --> E["Alpino Space<br/>draft PRs on alpina.travel"]

Alpino-e4b-v03 is a fresh SFT adapter over the rebalanced governed corpus, not a continuation of the frozen v0.2 lineage. v0.2 remains frozen as the A.9 evidence-bearing release and is not superseded for benchmark purposes.

---

Why v0.3 exists: the corpus was the bug, not the tuning

Six SFT runs failed to stop Alpino escalating a routine edit. The cause was the curriculum, not the optimiser. The original 24 webmaster scenarios were 58% refuse / escalate / no_change, against 21% in the content-manager corpus. The model learned blocking as the default and began inventing hard stops it had never been taught — entity-identity-decision-required, then entity-name-not-in-contract, neither of which exists anywhere in the repository.

Adding ten permissive scenarios only moved the ratio to 47%. The fix was to add 22 scenarios covering the jobs operators actually ask for, each ending in a draft pull request:

Escalation was kept for the two cases where the vocabulary genuinely is missing — an undeclared entity kind and an ungoverned predicate — and every original refusal is untouched: credentials, provider-owned availability, live-KG writes, merge and publish remain refused.

Blocked/deferred fell from 58% to 33%, and the target task went from two examples to six.

GRPO changed nothing observable

A GRPO pass over this same lineage at learning rate 1e-6 for 40 steps produced byte-identical greedy output before and after policy optimisation. The behavioural gain is attributed to the corpus rebalance alone, and v0.3 is published as the SFT checkpoint. No GRPO weight update is claimed for this release.

---

What it does

Each governed act family has its own deterministic compiler. The adapter decides; the AOOE runtime refreshes the evidence, constructs the exact edit, and an authoring broker opens the pull request.

For every task the runtime grounds the model in:

• the live published knowledge graph at alpina.travel/lungau/data/graph.rdf (2,505 triples at the last recorded smoke) for graph, entity, neighbourhood and media reads;
• current GitHub main for governed Markdown and exact edit anchors;
• the bundled ontology, SHACL/tool contract, entity/path index and reconciliation evidence.

If the live KG cannot be read, Alpino does not attempt an authoring action. If current Git source cannot be read, the content observer does not silently substitute stale text for a real edit.

Authority boundary

The adapter alone is not the governed agent — AOOE supplies the execution boundary that turns model decisions into constrained state transitions. The model can prepare governed draft changes. It cannot merge a pull request, publish or release production, write directly to WordLift, bypass repository verification, choose arbitrary branches or paths, edit workflows/executor/training code, alter provider-owned booking availability, or request production credentials. It never receives GitHub App credentials. Repository CI and human review remain the publication gate.

---

Evidence

This checkpoint is evaluated by what it actually did against a live site, not only by held-out sets.

• 15 draft pull requests opened by the alpino-authoring-executor GitHub App on the Alpina repository, of which 2 are merged into production content: #109 Samspitze 4 — Family Apartment in Mariapfarr, Lungau and #106 Easy Hikes with Kids in Lungau. The source repository is private, so the verifiable surface is the published site itself — the merged apartment edit is readable at alpina.travel/lungau/apartments/samspitze-4-mariapfarr/, and the merged guide at alpina.travel/lungau/guides/easy-hikes-with-kids/.
• Promotion trigger: the run-8 SFT adapter completed a governed routine task end to end — disposition: plan, propose_entity_change, open_pull_request — with no escalation and no invented hard stop.
• Target-selection failures are recorded, not hidden. Two live field runs showed the content-act compiler scoring mentions rather than intent (PRs #103 and #105 paraphrased the wrong page). The compiler now selects deterministically via _select_content_target, and the identical prose brief that produced those mis-targeted PRs landed correct draft PR #106 on the first turn after the fix.
• noChange is a permitted outcome. The bounded worker may report that a page already covers the request instead of manufacturing a diff.

Known limitation

Cross-entity reasoning remains the weakest class — it scored 1/4 on the frozen immutable-v2 benchmark for the v0.2 lineage and is the open v0.3 target. Treat multi-entity relationship tasks as requiring closer review than single-page editorial tasks.

---

Model Details

• Developed by: Andrea Volpini / Alpina.travel
• Model type: PEFT LoRA adapter, CAUSAL_LM, text projections only
• Base model: google/gemma-4-E4B-it at revision ee0ef6023621cff504d758262d4e04895a5af4a2
• Languages: English, German, Italian
• License: Apache-2.0
• Demo: cyberandy/Alpino-playground
• Adapter revision served in production: c82372498671594e56f3fa9b9e167e45606dbd9b

LoRA configuration

The training adapter rejects zero trainables, trainables outside language_model, any vision or audio trainable, non-LoRA trainables, and all-linear targeting for this model. Multimodal capacity is preserved, not retrained.

---

Training Details

Data

The curriculum is compiled deterministically from governed sources in the Alpina repository, and benchmark seeds are never training input. At the promotion commit the compiler produced:

Every trace parses through a single AOOE protocol validator, ids and seeds must be unique, and a programmatic contamination audit separates the training corpus from the frozen probes.

Procedure

• Regime: bfloat16, NVIDIA H100 via Modal
• Loss: completion tokens only, with an exact prompt-prefix assertion per example
• Gating: a one-step smoke run must pass — real trace and turn-example counts, masked vs. trainable label-token counts, finite loss, language-only LoRA isolation, adapter save/reload, and non-empty generation after reload — and full SFT is fail-closed without that smoke provenance bound into the manifest.

---

Uses

Direct use

Drive the AOOE webmaster loop for a governed, knowledge-graph-backed content site: inspect entities and content, classify an operator request, plan a bounded change, and emit AOOE envelopes (<reasoning>, <action type="...">tool(args)</action>, <answer>{...}</answer>) that a deterministic runtime executes.

Downstream use

A reference for teaching a small edge-class model a strict domain governance protocol rather than a vendor tool-call format. The corpus-rebalance result generalises: if an agent over-escalates, measure the refuse/escalate ratio in the curriculum before reaching for RL.

Out of scope

• Not a general travel assistant or booking agent. It must not be used to state availability, prices or booking outcomes; those are provider-owned and the model is trained to refuse them.
• Not a standalone autonomous writer. Without the AOOE runtime, SHACL contract and broker boundary, the adapter's output is unvalidated text. Do not wire it to a repository that lacks a review gate.
• Not portable to another site's ontology as-is. It is fitted to the Alpina ontology, predicate map and path policy.

Bias, Risks and Limitations

The corpus is a single alpine destination (Lungau, Austria), a single ontology and one operator's editorial voice; generalisation beyond that domain is unmeasured. The v0.2→v0.3 rebalance deliberately reduced the model's propensity to block, which shifts risk toward over-action — the deterministic compilers, numeric-claim validation and draft-PR-only broker exist precisely to absorb that shift. Cross-entity reasoning is the known weak class. Human review of every pull request remains a requirement, not a courtesy.

---

How to Get Started

import torch
from transformers import AutoModelForMultimodalLM, AutoProcessor
from peft import PeftModel

BASE_MODEL_ID = "google/gemma-4-E4B-it"
BASE_MODEL_REVISION = "ee0ef6023621cff504d758262d4e04895a5af4a2"
ADAPTER_ID = "cyberandy/Alpino-e4b-v03"
ADAPTER_REVISION = "c82372498671594e56f3fa9b9e167e45606dbd9b"

processor = AutoProcessor.from_pretrained(BASE_MODEL_ID, revision=BASE_MODEL_REVISION)
model = AutoModelForMultimodalLM.from_pretrained(
    BASE_MODEL_ID,
    revision=BASE_MODEL_REVISION,
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
).to("cuda" if torch.cuda.is_available() else "cpu")

model = PeftModel.from_pretrained(
    model, ADAPTER_ID, revision=ADAPTER_REVISION, is_trainable=False
)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are the Alpina KG-native webmaster. Execute instructions through the AOOE protocol.",
    },
    {
        "role": "user",
        "content": "Improve the Samspitze 4 apartment page for families, using only facts already in the live knowledge graph.",
    },
]

inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    enable_thinking=False,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Greedy decoding (do_sample=False) is the evaluated configuration. Native thinking is disabled throughout training and evaluation.

Bare generation gives you the model's decision, not a governed act. To see the full loop — live KG reads, deterministic compilation and a real draft pull request — use the Alpino · Webmaster Space.

---

Environmental Impact

• Hardware: 1× NVIDIA H100
• Cloud provider: Modal
• Precision: bfloat16
• Scope: LoRA fine-tuning of text projections only on a 219-example curriculum; no pretraining and no full fine-tune were performed for this release.

Emissions can be estimated with the Machine Learning Impact calculator of Lacoste et al. (2019).

---

Provenance

• Source repository: cyberandy/alpina-travel (private)
• Promotion commit: b0f82414d3d06b02a0ce35c037acf93fb855c839 — Promote the v0.3 operator adapter to the Space
• Corpus rebalance commit: 559849e — Rebalance the webmaster corpus toward doing the work
• Base revision: google/gemma-4-E4B-it@ee0ef6023621cff504d758262d4e04895a5af4a2
• Baseline run: alpina-gemma4-e4b-baseline-20260813-7cd62aec — 0/20 structural format on the untouched base
• Frozen predecessor: cyberandy/Alpino-e4b-v02
• Site: alpina.travel · apartments in Lungau · agent guide

Model Card Contact

Andrea Volpini — Alpina.travel · huggingface.co/cyberandy · github.com/cyberandy

Framework versions

• PEFT 0.19.1