Instructions to use AIHeikki/lisa-gemma4-e4b-visual-auditor with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries
PEFT
How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Use Docker

docker model run hf.co/AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AIHeikki/lisa-gemma4-e4b-visual-auditor"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AIHeikki/lisa-gemma4-e4b-visual-auditor",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Ollama
How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with Ollama:
```
ollama run hf.co/AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M
```
Unsloth Desktop

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with Docker Model Runner:
```
docker model run hf.co/AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M
```

Lemonade

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Run and chat with the model

lemonade run user.lisa-gemma4-e4b-visual-auditor-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AIHeikki/lisa-gemma4-e4b-visual-auditor with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "AIHeikki/lisa-gemma4-e4b-visual-auditor:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Lisa — Gemma-4-E4B-it Multimodal Visual Slide & Brand Auditor (`Exp 5B`)

Lisa (lisa-gemma4-e4b-visual-auditor) is our multimodal vision-language forensic auditor (Exp 5B): a google/gemma-4-E4B-it model jointly fine-tuned across both Vision Encoder layers and Language Decoder layers (r = 16, alpha = 16 BF16 LoRA) and exported as a standalone 4.97 GiB Q4_K_M GGUF + 945.6 MiB BF16 mmproj Vision Projector GGUF for Executive Slide Forensics, Chart Geometry Auditing, Cross-Slide Contradiction Detection, and Brand Typography/Palette Compliance.

It is part of the Mona, Lisa & Fable Trilogy (Exp 5A–5C):

AIHeikki/mona-qwen3.5-4b-epistemic-sparring (Exp 5A): Textual & RAG Epistemic Discipline (Qwen3.5-4B).
AIHeikki/lisa-gemma4-e4b-visual-auditor (Exp 5B, this repository): Multimodal Slide & Visual Brand Forensics (Gemma-4-E4B-it).
AIHeikki/fable-mistral-small-4-119b-telco-lora (Exp 5C): Frontier-scale (119B MoE) 5G SA Telco Architect & NOC Root-Cause Commander.

1. What Lisa Detects in Slide Decks & Technical Diagrams

Base vision-language models act as polite OCR summarizers: if a bar chart label says "48%" while the bar is drawn at 85% of the plot height, or if a slide deck violates corporate hex palettes and minimum font sizes, vanilla VLMs praise the slide's layout.

Lisa was trained on programmatically rendered high-resolution (1600x900) single-slide and multi-slide pairs (render_deck_dataset.py) with ground-truth pixel bounding boxes and geometric distortions across 5 forensic dimensions:

Chart Lie & Truncated Axis Detection: Catches bar-height vs. data-label mismatches, truncated non-zero baselines that exaggerate slopes by 3x–5x, and dual-axis scale manipulation.
Cross-Slide Contradiction Forensics: Compares multi-slide pairs (e.g., Executive Summary vs. Appendix Financial Table) to flag conflicting ARR, margin, headcount, or timeline numbers.
Brand Palette & Contrast Compliance: Flags off-palette accent colors, low-contrast text-on-background failures, and unauthorized template modifications.
Typography & Information Density: Detects overcrowding, tiny footnote sizes, and broken visual hierarchy.
Calibrated Clean-Slide Pass Rate: Passes genuinely compliant slides without inventing false-positive critiques.

2. Held-Out Multimodal Benchmark (`N = 50` Unseen Rendered Slide Decks)

Evaluation Metric (`N = 50` Unseen Decks across 5 Categories)	Base `google/gemma-4-E4B-it`	Lisa `Gemma-4-E4B` (`Exp 5B`)	Absolute Gain
Overall Strict Pass Rate (verdict + defect + severity + spatial grounding)	`32.0%`	`92.0%`	`+60.0 pp`
1. Chart Lie & Axis Distortion Detection (`N = 10`)	`20.0%`	`90.0%`	`+70.0 pp`
2. Cross-Slide Numerical Contradiction (`N = 10`)	`30.0%`	`90.0%`	`+60.0 pp`
3. Brand Palette & Contrast Compliance (`N = 10`)	`10.0%`	`100.0%`	`+90.0 pp`
4. Typography & Density Auditing (`N = 10`)	`20.0%`	`80.0%`	`+60.0 pp`
5. Clean Compliant Slide Calibration (`N = 10`)	`80.0%`	`100.0%`	`+20.0 pp`
Spatial Quadrant & Coordinate Grounding Recall	`28.5%`	`93.5%`	`+65.0 pp`

3. Repository Layout

.
├── README.md                                  # This Model Card
├── lisa-gemma4-e4b-Q4_K_M.gguf                # 4.97 GiB Q4_K_M language + multimodal backbone GGUF
├── mmproj-lisa-gemma4-e4b-BF16.gguf           # 945.6 MiB BF16 Vision Projector GGUF (required for image input in llama.cpp)
├── Modelfile                                  # Ollama Modelfile
├── run_lisa_server.sh                         # One-command llama-server launch script with mmproj attached
├── lora_adapter/                              # BF16 PEFT LoRA adapter (r=16, alpha=16 across Vision + Language layers, 186.2 MB)
│   ├── adapter_model.safetensors
│   ├── adapter_config.json
│   ├── chat_template.jinja
│   ├── processor_config.json
│   ├── tokenizer.json
│   └── tokenizer_config.json
└── reports/
    ├── EVALUATION_REPORT_EXP5B.md             # Full 50-deck multimodal benchmark report
    ├── MONA_AND_LISA_EXPERIMENT_STORY.md      # Narrative overview of Exp 5A (Mona) & Exp 5B (Lisa)
    └── LISA_SOVEREIGN_AI_SDLC_AUDIT.md        # Real-world slide-by-slide visual audit example

4. Quickstart (`llama.cpp` Vision Server)

llama-server \
  -m lisa-gemma4-e4b-Q4_K_M.gguf \
  --mmproj mmproj-lisa-gemma4-e4b-BF16.gguf \
  -c 8192 -ngl 99 --temp 0.1 --port 8080

5. License

Released under the Apache 2.0 License, matching the upstream google/gemma-4-E4B-it license.

Downloads last month: 8

GGUF

Model size

8B params

Architecture

gemma4

Hardware compatibility

4-bit

Model tree for AIHeikki/lisa-gemma4-e4b-visual-auditor

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Adapter

(488)

this model

Lisa — Gemma-4-E4B-it Multimodal Visual Slide & Brand Auditor (Exp 5B)

1. What Lisa Detects in Slide Decks & Technical Diagrams

2. Held-Out Multimodal Benchmark (N = 50 Unseen Rendered Slide Decks)

3. Repository Layout

4. Quickstart (llama.cpp Vision Server)

5. License

Model tree for AIHeikki/lisa-gemma4-e4b-visual-auditor

Lisa — Gemma-4-E4B-it Multimodal Visual Slide & Brand Auditor (`Exp 5B`)

2. Held-Out Multimodal Benchmark (`N = 50` Unseen Rendered Slide Decks)

4. Quickstart (`llama.cpp` Vision Server)