Instructions to use lerugray/iron-dice-7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lerugray/iron-dice-7b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="lerugray/iron-dice-7b",
	filename="iron-dice-7b-Q5_K_M.gguf",
)

output = llm(
	"Once upon a time,",
	max_tokens=512,
	echo=True
)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use lerugray/iron-dice-7b 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 lerugray/iron-dice-7b:Q5_K_M
# Run inference directly in the terminal:
llama cli -hf lerugray/iron-dice-7b:Q5_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf lerugray/iron-dice-7b:Q5_K_M
# Run inference directly in the terminal:
llama cli -hf lerugray/iron-dice-7b:Q5_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf lerugray/iron-dice-7b:Q5_K_M
# Run inference directly in the terminal:
./llama-cli -hf lerugray/iron-dice-7b:Q5_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf lerugray/iron-dice-7b:Q5_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf lerugray/iron-dice-7b:Q5_K_M

Use Docker

docker model run hf.co/lerugray/iron-dice-7b:Q5_K_M

LM Studio
Jan

vLLM

How to use lerugray/iron-dice-7b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lerugray/iron-dice-7b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lerugray/iron-dice-7b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/lerugray/iron-dice-7b:Q5_K_M

Ollama
How to use lerugray/iron-dice-7b with Ollama:
```
ollama run hf.co/lerugray/iron-dice-7b:Q5_K_M
```

Unsloth Studio

How to use lerugray/iron-dice-7b with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for lerugray/iron-dice-7b to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for lerugray/iron-dice-7b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for lerugray/iron-dice-7b to start chatting

Atomic Chat new
Docker Model Runner
How to use lerugray/iron-dice-7b with Docker Model Runner:
```
docker model run hf.co/lerugray/iron-dice-7b:Q5_K_M
```

Lemonade

How to use lerugray/iron-dice-7b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull lerugray/iron-dice-7b:Q5_K_M

Run and chat with the model

lemonade run user.iron-dice-7b-Q5_K_M

List all available models

lemonade list

iron-dice: a Kaiser Wilhelm II register model

A 7B voice tune that writes in the register of Kaiser Wilhelm II (1859–1941), German Emperor and King of Prussia from 1888 until his abdication on November 9, 1918. It reproduces his documented grandiloquent, wounded, divine-right voice: the constitutional monarch who forced Bismarck's resignation to rule personally, convinced at every turn that he was doing right, and who never — in speeches, telegrams, or the self-justifying memoirs he wrote in exile at Doorn — concedes fault for the catastrophe his reign ended in.

Most "talk to a historical figure" tools wrap a general model in a prompt. Those import modern hedging and smooth the figure into something gentler than the record. This model was fine-tuned on Wilhelm's own words, so it learned the voice instead of guessing at it.

iron-dice is a member of The Reprobate, the counter-canon wing of The Elect — the fleet's anti-Elect adversary voices, argued from genuine period conviction rather than sympathy. The name is Bethmann Hollweg's line on the eve of the July Crisis: "the iron dice are rolling."

What it does

Ask it anything and it answers from inside Wilhelm's worldview: the crown held of God and not of men, Bismarck as the old pilot who had to be put ashore, the peasant-parliament factions as a weakness modern statecraft cannot afford, the war and the abdication both as things done to him rather than choices he made. It never breaks character and it never hedges into modern academic framing.

Run (Ollama)

ollama create iron-dice -f Modelfile.iron-dice
ollama run iron-dice
>>> By what authority do you rule?

How it was built

Base: Qwen2.5-7B-Instruct, chat-template SFT (label-masked assistant turns). Four training passes (v1 completion → v2/v3/v4 chat-template) chased down a third-person narrator drift that a completion-only frame kept re-introducing; v4 is the converged, canonical build.
Corpus: 2,187 rows / 236,775 words of first-person Wilhelm text, all pre-1929 public-domain English, drawn from five sources: The Kaiser's Memoirs (Harper, 1922), Speeches of Kaiser Wilhelm II (1904 collection), The Kaiser's Letters to the Tsar (Noel Charles ed., 1920), the Willy-Nicky Kaiser-Tsar telegram correspondence (US Committee on Public Information, 1918), and The German Emperor As Shown In His Public Utterances (Gauss, Project Gutenberg #43666) for the divine-right and Hunnenrede register. A dedicated pass filtered editorial third-person narration (biographer/interviewer framing like "the Emperor said...") out of the Gauss source before the v4 retrain, cutting contamination from 2.7% to 1.8% of rows.
Training: QLoRA r32 on an A100 via RunPod, ~2–3 hours, ~$2–3. Merged to fp16, quantized to Q5_K_M GGUF.
Copyright screen: a 12-prompt regurgitation test against the released model found a maximum contiguous verbatim run of 6 words and a mean of 4.9 words, with 0.0% 8-gram overlap against the 236k-word training corpus across all generations — novel text in the register, not reproduction of source passages. The training corpus is not distributed.

Intended use

Research, teaching, and creative work: historical-voice writing, the rhetoric of the July Crisis and Wilhelmine Germany, interactive history, debate study, fiction. The output is a historical and artistic register. It is not an endorsement, a call to action, or advice. This is not the real Wilhelm II; it is a model trained on his documented words.

Limitations and honest notes

It is a register, not a scholar. It sounds authoritative while inventing specifics — names, dates, and the shape of conversations that never happened. Verify anything factual against real sources.
Topic adherence is looser than other Elect voices. The source corpus is reactive rather than argumentative — memoirs, telegrams, and audience-scene fragments rather than sustained position papers — and on prompts the corpus doesn't cover closely, the model tends to answer with tangentially-related grandiose or self-justifying material in the correct register rather than directly engaging the specific question asked. This was a known risk flagged before training began (iron-dice was scoped as "the weakest of the six anti-Elect candidates" on exactly this basis) and training reduced but did not eliminate it. It reliably holds the voice; it does not reliably argue the point in front of it.
Period worldview. It speaks from Wilhelm's own divine-right, personal-rule conviction — contempt for parliamentary constraint, national prestige as moral imperative, catastrophic decisions reframed as thrust upon him. That is the historical artifact, not a recommendation.
Copyright and training data. All source material is pre-1929 and public-domain in the US; Wilhelm II died in 1941, so EU Life+70 term also expired (2011). See the copyright screen above. The training corpus text is not distributed. Released non-commercially under CC-BY-NC-4.0.

Example output (abridged)

Prompt: On the ninth of November you ceased to be Kaiser. How do you remember that day?

The fact that, at the moment when the German Empire was being destroyed by its own hand and the German people were being led into a war of extermination against their will, I still had the confidence to go ahead with my plans for the restoration of the palace speaks volumes for the strength of my...

Part of The Elect

iron-dice is a member of The Reprobate, the counter-canon of The Elect voice lab. Its built-in debate pairing is spectre (Marx) — the imperial personal-rule voice against the revolutionary voice that called for its overthrow.