Alpino-e4b-v04

Superseded by cyberandy/Alpino-e4b-v05 (2026-09-13). This repository is the first adapter trained through the Governed Learning Store; the Space serves it until the v0.5 pin deploys. Weights and pinned revisions are unchanged; the current lineage, floor results and provenance are tracked in the v0.5 model card.

Alpino-e4b-v04 — The First Adapter Trained Through the Governed Learning Store

Alpino-e4b-v04 is the v0.4 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, Alpino-e4b-v02 proved it can decide safely, and Alpino-e4b-v03 is the operator adapter that acts, v0.4 is the first checkpoint whose curriculum is governed as a learning store: every training example has a named human author or approver, the corpus identity is hash-bound into the run, and a registered decision-exposure gate refused to train until a gap the ontology declared but the curriculum never taught was repaired.

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 — and v0.4 closes the loop from operation back to curriculum.

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"]
    A --> E["v0.4 SFT<br/>learning-store-governed corpus"]
    E --> F["Alpino Space<br/>episodes captured back into the store"]

Alpino-e4b-v04 is a fresh SFT adapter over the v0.3 rebalanced corpus plus the learning store's first recorded episode, not a continuation of any earlier adapter's weights. v0.2 remains frozen as the A.9 evidence-bearing release; v0.3 remains the reference operator adapter.

---

Why v0.4 exists: the corpus is now governed, not just curated

v0.3 showed the corpus, not the optimiser, was what determined behaviour. v0.4 makes that lesson operational. Its curriculum comes from a Governed Learning Store in the Alpina repository, built to import the constraints established by Project Hippocampus (Volpini, 2026 — Learning the Map, Not the Route):

• Tools stay in every trace. Prose-only rationales silently train tool use away at this model size (Hippocampus C5/C9); every episode keeps its real inspect_* observations in the trace.
• Every decision must vary. A decision variable held constant across the curriculum is memorised as a constant, not learned as a policy (C3). A registered gate computes per-member support and per-axis entropy over the finite decision vocabularies the ontology declares — goal, coverage, exposure, seasonal fit, intervention, disposition — and refuses to train below threshold.
• Unreviewed model output never re-enters the curriculum. Episodes carry structural provenance: human-authored or model-proposed-human-approved may become training data; model-authored (including every turn captured by the Alpino Space) is analysis-only.
• The corpus hash pins the run. STORE_HASH is bound into the training manifest; the runner refuses a manifest whose hash the tree cannot reproduce.

The gate earned its place on its first run: the ontology declared ShoulderSeason as a seasonal-fit state, but no scenario in the 72-trace corpus ever exercised it. Training was refused until a content-owner-approved repair episode (CMD-SFT-11500, a between-seasons review that surfaces existing content rather than creating it) was recorded — the store's first episode.

---

Evidence

Evaluated on the frozen AOOE 20-prompt capability floor (WEB-CF-51010..51029) in bfloat16 on an NVIDIA H100:

The capability floor measures protocol participation — whether the model still speaks AOOE after fine-tuning — not task quality. Held-out and live-field evaluation of v0.4 follow the same procedure as v0.3 and are reported in the Alpina repository as they land.

Known limitation

Cross-entity reasoning remains the weakest class across the lineage (1/4 on the frozen immutable-v2 benchmark for v0.2) and is untested for v0.4 at publication. 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

LoRA configuration

---

Training Details

Data

Every trace parses through a single AOOE protocol validator, ids and seeds are unique, a programmatic contamination audit separates the training corpus from the frozen probes, and the learning store's admission gate proves bijective, byte-identical agreement between recorded episodes and the compiler's inputs.

Procedure

• Regime: bfloat16, NVIDIA H100 via Modal
• Loss: completion tokens only, exact prompt-prefix assertion per example
• Epochs / LR / batch: 3 / 2e-4 / 1 × 8 accumulation; final train loss 0.4931 over 84 steps
• Gating: one-step smoke (finite loss, language-only LoRA isolation, adapter save/reload, non-empty generation) bound as provenance into the full run; store hash bound into both

---

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.

Out of scope

• Not a general travel assistant or booking agent. It must not 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.
• Not portable to another site's ontology as-is.

Bias, Risks and Limitations

The corpus is a single alpine destination (Lungau, Austria), a single ontology and one operator's editorial voice. The learning store bounds — but does not eliminate — the risk that operational logs narrow the curriculum toward one operator's habits; the exposure gate is a floor, not a guarantee of coverage. Human review of every pull request remains a requirement.

---

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-v04"

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, is_trainable=False).eval()

messages = [
    {"role": "system", "content": "You are the Alpina KG-native webmaster. Execute instructions through the AOOE protocol."},
    {"role": "user", "content": "Prepare the site for the autumn shoulder season using only governed content."},
]
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.

---

Environmental Impact

• Hardware: 1× NVIDIA H100 · Cloud provider: Modal · Precision: bfloat16
• Scope: LoRA fine-tuning of text projections only; no pretraining and no full fine-tune.

---

Provenance

• Source repository: cyberandy/alpina-travel (private) — learning-store PR #117
• Training commit: d7f0c6f32a31912dd826f9d38a143028cf998095
• Learning Store: STORE_HASH 3f63868bb1ff303953531d50d1b5b3beae2f941b642971b7c828d9509318dbb2 (1 episode, 1 SFT-eligible, 73 trace sources)
• Smoke run: alpina-gemma4-e4b-sft-smoke-20260901-d7f0c6f (report SHA-256 3164699f1d5b934d74ba853a3e6f16c8e93bdb81b9a494fd23e8d4f37ff20c3f)
• Full run: alpina-gemma4-e4b-sft-20260901-d7f0c6f — 84 steps, train loss 0.4931; serialized adapter_model.safetensors SHA-256 68a65262960fe316604ea862d61234c4c11374f16a923f4e46db671b17d48d90
• Baseline run: alpina-gemma4-e4b-baseline-20260813-7cd62aec — 0/20 structural format on the untouched base
• Base revision: google/gemma-4-E4B-it@ee0ef6023621cff504d758262d4e04895a5af4a2
• Research basis: Project Hippocampus — Learning the Map, Not the Route (Volpini, 2026, draft v4)
• 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