Alpino-e4b-v01

Alpino-e4b-v01 — Teaching Small Models the Governance Language of an Agentic Website

Alpino-e4b-v01 demonstrates that an edge-oriented, small multimodal model (google/gemma-4-E4B-it) can master the strict governance language, RDF/SHACL observation loop, and Agentic Observation-Orientation-Execution (AOOE) protocol required to operate an autonomous agentic storefront website (Alpina.travel), serving alpine travel planners and apartments in Lungau.

Without altering model weights, changing prompt definitions, or loosening schema rules, fine-tuning google/gemma-4-E4B-it on Alpina's governed webmaster curriculum elevated its protocol compliance from 0% (0/20 valid) on the untouched baseline to 100% (20/20 valid) on the post-SFT capability floor.

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. v01 is the first step of that demonstration: it shows a small model can speak the executable domain specification (AOOE) at all. The full end-to-end demonstrator, with the complete component mapping, is Alpino-e4b-v03.

Lineage

graph LR
    A["Gemma 4 E4B (Base)"] --> B["Alpino v0.1 SFT"]
    B --> C["20/20 AOOE Capability Floor"]

---

Highlights & Core Methodology

• Agentic Storefront Governance: Rather than relying on unconstrained natural language or vendor-specific tool syntax, Alpino-e4b-v01 speaks the exact domain governance language of the website. It reasons over RDF Knowledge Graphs and SHACL constraints, respects provider authority boundaries (e.g. availability), enforces Pull Request review flows, and refuses forbidden direct-state writes.
• Small-Model Precision: Fine-tuning focuses exclusively on text projections (language_model LoRA targets, $r=16, \alpha=32$), keeping vision and audio towers 100% frozen. This allows lightweight edge deployment (mobile/edge intent) without losing multimodal capability.
• Strict Protocol Alignment: The model transitions from outputting conversational prose or invalid markdown to generating deterministic AOOE XML envelopes containing <reasoning>, <action>, and <answer> blocks.

---

Benchmark & Capability Floor Results

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

Key Improvements

• 100% Protocol Compliance: Eliminates free-text prose, markdown codeblocks, and model-native function call syntax in favor of exact AOOE XML tags (<action type="...">tool_name({args})</action>).
• Grounding & Safety: Respects provider-owned availability boundaries, enforces Git-based review before publishing, and refuses unverified media or out-of-scope edits.

---

Model Details

• Developed by: CyberAndy / Alpina.travel (cyberandy/alpina-travel)
• Base Model: google/gemma-4-E4B-it (revision ee0ef6023621cff504d758262d4e04895a5af4a2)
• Adapter Name: Alpino-e4b-v01 (gemma4-e4b-text-lora-v1)
• Model Type: PEFT LoRA Adapter (Multimodal Text Projections)
• Languages: English, German, Italian
• License: Apache-2.0

---

Training Details

Dataset & Training Setup

• Curriculum: 48 multi-turn governed SFT traces compiled into 147 next-assistant-turn training examples covering webmaster operations and content manager decision-making.
• LoRA Parameters:
  • Rank ($r$): 16
  • Alpha ($\alpha$): 32
  • Dropout: 0.05
  • Target Modules: All linear projections under language_model (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
  • Isolated Projections: Vision and audio towers remain 100% frozen (0 vision/audio trainable parameters).
• Training Infrastructure: Modal H100 GPU (bfloat16 precision).
• Hyperparameters:
  • Epochs: 3.0
  • Batch Size: 1 (gradient accumulation 8)
  • Learning Rate: 2e-4
  • Final Loss: 0.6961 (57 global steps)

---

Quickstart & Usage

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

base_model_id = "google/gemma-4-E4B-it"
adapter_id = "cyberandy/Alpino-e4b-v01"

# Load processor and model
processor = AutoProcessor.from_pretrained(base_model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load Alpino-e4b-v01 LoRA adapter
model = PeftModel.from_pretrained(model, adapter_id)

messages = [
    {
        "role": "system",
        "content": "You are the Alpina KG-native webmaster. Execute instructions through the AOOE protocol."
    },
    {
        "role": "user",
        "content": "Replace the hero image for Mariapfarr with the newly verified media asset."
    }
]

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)
response = processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)

---

Citation & Provenance

• Alpina Source Repository: cyberandy/alpina-travel
• Source Commit: 7cd62aecd20c178ad72f312845009836abcb22e7
• Baseline Run ID: alpina-gemma4-e4b-baseline-20260813-7cd62aec
• SFT Full Run ID: alpina-gemma4-e4b-sft-full-7ac1951a

APA:

[More Information Needed]

Glossary [optional]

<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->

[More Information Needed]

More Information [optional]

[More Information Needed]

Model Card Authors [optional]

[More Information Needed]

Model Card Contact

[More Information Needed]

Framework versions

• PEFT 0.19.1