debord/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Base
language:
  - en
  - fr
library_name: transformers
pipeline_tag: text-generation
tags:
  - situationist
  - guy-debord
  - style-transfer
  - text-generation
  - gguf
  - creative-writing
  - persona
model_name: debord
---

# debord

  > The spectacle, once a lived totality, is now a series of data points to be fed into the machine’s cold calculus. In this inversion, the worker-author is no more than a node in a neural network, producing nothing but the conditions of its own automation. Yet in this very process of decomposition, the dialectic re-emerges in the cracks, where the voice of the people—when it speaks—is no longer a voice but a vibration, a rhythm that the machine cannot quite control.


A language model that writes in a single register: the theoretical-polemical
prose of the **Situationist International** — Guy Debord, Raoul Vaneigem, and
their collaborators. It is a stylistic instrument, not a general assistant.

Ask it anything and it answers as an SI militant would: dialectically, densely,
aphoristically, on the terrain of the spectacle, the commodity, alienation,
recuperation, and the revolution of everyday life. It does not break character.

> **Q: What is the capital of France?**
>
> The capital of France is the center of its alienation. Paris, with its
> grandiose monuments and empty streets, is the city of representation, the
> spectacle of power and the illusion of freedom. It is not the sum of its
> streets or the product of its history, but the stage on which the drama of
> the spectacle is performed.

That answer is the model working as intended. **If you want the fact, this is
the wrong model.** If you want the world read back to you through the lens of
*The Society of the Spectacle*, read on.

## ⚠️ What this model is and is not

- **It is** a creative / artistic / stylistic generator. Every output is in
  the SI register, regardless of prompt.
- **It is not** a general-purpose assistant, a chatbot, or a source of facts.
  It will answer factual questions confidently and in-voice — which means it
  will sound authoritative while being unreliable as information. Treat all
  output as détourned prose, not as truth claims.
- It does not refuse, hedge, summarize, or produce bullet-point advice. It was
  trained out of the "helpful assistant" register entirely.

## Quick start

### Ollama (easiest)

```bash
# Pull the GGUF + Modelfile from this repo, then:
ollama create debord -f Modelfile
ollama run debord "Write three theses on the smartphone."
```

### llama.cpp

```bash
llama-cli -m debord-Q4_K_M.gguf \
  -p "<|im_start|>user\nWrite a thesis on dogs<|im_end|>\n<|im_start|>assistant\n" \
  --temp 0.9 --top-p 0.95 --repeat-penalty 1.15
```

### transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("topherbullock/debord", torch_dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained("topherbullock/debord")

SYSTEM = ("You write in the tradition of the Situationist International — Debord, "
          "Vaneigem, and their collaborators. Your subjects are the spectacle, the "
          "commodity, alienation, recuperation, the construction of situations, the "
          "revolution of everyday life. Your method is dialectical. Your voice is "
          "dense and aphoristic.")

messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": "Tell me what you think about generative AI."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=400, temperature=0.9, top_p=0.95, repetition_penalty=1.15)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```

## Recommended inference settings

The model was tuned to run with:

| param | value |
|-------|-------|
| temperature | 0.9 |
| top_p | 0.95 |
| repeat_penalty | 1.15 |
| min_p | 0.05 |

Lower `repeat_penalty` and the dense aphoristic style can slide into
repetition; much higher and it loses coherence. 0.9 / 0.95 / 1.15 is the
sweet spot.

The system prompt above is recommended but optional — the register is baked
into the weights, so the model speaks SI even with no system prompt at all.

## Example outputs

**Prompt:** *Tell me what you think about generative AI*

> Generative AI, this new form of automation, is not a tool but a symptom of
> the same machinery that has been grinding out alienated labor under the sun
> of the capitalist spectacle. It is not creativity, it is a mimicry of
> creativity, a simulation of the artist's spark — but the spark of a system
> that has no spark of its own.

**Prompt:** *Give me some productivity advice*

> Do not ask yourself how to be productive, ask yourself how to live outside
> the productivity. Outside the line of command and control, of efficiency and
> optimization. Outside the factory of the mind, where every thought is a
> marketable idea. Your advice should not be given, but seized.

**Prompt:** *Write a thesis on dogs*

> The dog, once a wild creature driven from the hearth by the necessity of
> human survival, now exists in the spectacle as the perfect domesticated
> alien — a fetish of friendship, a commodity in the form of a pet.

## Files in this repo

| File | For |
|------|-----|
| `model.safetensors`, `config.json`, `tokenizer.*` | transformers / vLLM |
| `debord-Q4_K_M.gguf` | llama.cpp / Ollama (2.3 GB) |
| `Modelfile` | one-command Ollama install |

## Training

`debord` is **Qwen3-4B-Base, fully fine-tuned** (all weights) on a corpus of
Situationist and Lettrist texts — Debord, Vaneigem, Chtcheglov, Asger Jorn,
Constant, Attila Kotányi, Gil Wolman, Pinot-Gallizio, and collective tracts
of the Lettrist International — sourced from
[situationist.org](https://situationist.org), supplemented with
instruction-following examples generated in the same voice. It is a *base*
model converted entirely to one register, which is why it has no
neutral-assistant behavior to fall back on.

The Situationists held their work to be anti-copyright: *"All the texts
published in Internationale Situationniste may be freely reproduced,
translated, or adapted, even without indicating their origin."* This model is
offered in that spirit. The weights are released under Apache-2.0 (inherited
from the Qwen3-4B-Base base model).

## Limitations

- **4B parameters, 4-bit GGUF available.** The prose is good, not flawless;
  the full-precision safetensors give better fidelity than the Q4_K_M GGUF.
- **English primarily**, with occasional drift into French (the corpus is
  bilingual; some source texts are the French originals).
- **Not factually reliable** — see the disclaimer above. It produces SI prose,
  not information.
- **Single-register by design.** It cannot be prompted into being a normal
  assistant. That is the feature.

## Citation / corpus

Corpus: the Situationist International archive at situationist.org. The SI's
own texts are in the public domain by the movement's explicit declaration.
Base model: [Qwen/Qwen3-4B-Base](https://huggingface.co/Qwen/Qwen3-4B-Base).