Wolfgang-LM (75M) - Instruct (Goethe Persona)
Wolfgang-LM-Instruct is the fine-tuned version of Wolfgang-LM-Base. It has been trained to embody the persona of Johann Wolfgang von Goethe, capable of answering modern queries in the linguistic style of Weimar Classicism.
⚠️ Research Preview: This model is a research artifact. While it attempts to simulate Goethe's views, it generates fictional content and should not be treated as a historical authority.
Model Details
- Model Type: Decoder-Only Transformer (Llama-style architecture)
- Parameters: ~75M
- Language(s): German (Historical: 1750-1832)
- License: MIT License
- Finetuned from: Wolfgang-LM-75M-Base
Uses
Intended Use
- Roleplay/Chat: Simulating conversations with a Goethe-like persona.
- Educational Engagement: Making 19th-century language accessible through interactive dialogue.
- Style Transfer: Reformulating modern concepts (e.g., "Artificial Intelligence") into historical German ("eine künstliche Seele").
Out-of-Scope Use
- Factual Accuracy: The model prioritizes style and persona over factual correctness.
- Modern Assistance: Do not use for medical, legal, financial, technical, or any other form of professional advice. The model will likely answer via metaphors or refusal.
Training Data
The model was fine-tuned on a mix of historical and synthetic dialogue data:
- Conversations with Goethe (Gespräche mit Goethe)
- Source: Johann Peter Eckermann's transcripts (DTA).
- Nature: Authentic historical dialogues used to capture Goethe's voice and mannerisms.
- Synthetic Instruction Tuning Data
- Source: Generated via Gemini 2.5 Flash.
- Nature: ~4,500 pairs of modern user queries and Goethe-styled responses. Includes small talk, identity questions, and refusals of toxic/modern requests.
How to Get Started
The model uses a specific prompt format with special tokens: <|user|>, <|assistant|>, <|system|>, and <|endoftext|>.
[!IMPORTANT] Critical Configuration:
- System Prompt: You SHOULD use the exact system prompt:
"Du bist Johann Wolfgang von Goethe.". The model was fine-tuned exclusively with this identity; changing it might break the persona.- Inference Settings: The sampling parameters (Temperature 0.5, Top-P 0.8, Top-K 15, Repetition Penalty 1.1) are tuned for stability. Deviating from these settings may cause the model to generate nonsensical or repetitive text.
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "dominik-reiner/wolfgang-lm-75M-instruct-v1"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# 1. Define the Chat
chat = [
{"role": "system", "content": "Du bist Johann Wolfgang von Goethe."},
{"role": "user", "content": "Was hältst du von der modernen Technik?"}
]
# 2. Format the Prompt manually (or use apply_chat_template if configured)
prompt = ""
for msg in chat:
prompt += f"<|{msg['role']}|>\n{msg['content']}\n<|endoftext|>"
prompt += "<|assistant|>\n"
inputs = tokenizer(prompt, return_tensors="pt")
# 3. Generate
outputs = model.generate(
**inputs,
max_new_tokens=100,
temperature=0.5,
top_p=0.8,
top_k=15,
repetition_penalty=1.1,
do_sample=True
)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))
Evaluation
The model was evaluated on a held-out validation set of synthetic conversations.
- Validation Loss: ~2.85
- Perplexity: ~17.26
Performance & Limitations
- Size Constraint: At 75M parameters, the model's reasoning capabilities are limited. It relies heavily on surface-level style imitation.
- Hallucinations: It may invent poems, citations, or historical events that never happened.
- Safety: The model includes safety training to refuse harmful requests, often citing "propriety" or "dignity" as the reason.
Resources
- Demo: Wolfgang-LM Play (HF Space)
- GitHub Repository: dominik-reiner/wolfgang-lm
- Architecture: Model Architecture
- Data Sources: Data Documentation
- Synthetic Data: Synthetic Data Details