README.md

---
language:
  - en
license: cc-by-nc-4.0
tags:
  - text-generation
  - history
  - ancient-greece
  - voice-model
  - ollama
  - elect
base_model: Qwen/Qwen2.5-7B-Instruct
---

# chaeronea: Philip II of Macedon (382–336 BC)

> **PLAUSIBLE PHILIP, a constructed voice-model.** Philip II left almost no first-person
> text. This model is trained first on his real recorded words, then on a plausible voice
> extrapolated from the secondary record. It is honest about the difference. Do not treat
> its outputs as historical quotations.

Part of the **Elect** voice-model series: historical figures as runnable instruments.

## What this is

A fine-tuned Qwen2.5-7B-Instruct that targets the documented register of Philip II of
Macedon: dry, ironic, pragmatic. Philip reasons from interest rather than principle. He
does not moralize or posture. His wit is the wit of a man who has bribed his way into more
cities than he has stormed, and who finds this more interesting than shameful.

The contrast with the other Elect voices is deliberate. Philip is the realpolitik voice
where Müntzer is apocalyptic fury, Marx is systemic critique, Joan is divine certainty.
Philip has no system. He has a ledger.

## The honest thinness

Philip's surviving first-person corpus is the thinnest of any figure in the series, thinner
even than Joan of Arc. What survives is a scatter of recorded speech: roughly thirty
apophthegms preserved by Plutarch, and one diplomatic letter. The model is built in two
layers, and the distinction is the whole point.

**T1, real recorded speech.** Plutarch, *Sayings of Kings and Commanders* (Apophthegmata
Regum, ch. 26), Babbitt translation (Loeb, 1931); and Demosthenes 12, *Philip's Letter*,
Vince & Vince translation (Loeb, 1926). Both are US public domain.

**T2, a constructed plausible Philip.** A separate pass extrapolates his voice from the
secondary record (Diodorus Siculus Book 16; Justin, *Epitome* Books 7 to 9), generated in
persona by an off-platform model and register-checked. This layer is constructed, not
recorded, and is labelled as such everywhere this model appears.

This is the closest approximation of Philip's register the record permits, labelled
honestly. It is not pretending to be Philip's ghost.

## Example exchanges

Real outputs from this model:

> **Is there any fortress that cannot be taken?**
> Yes, the tomb.

> **Athens calls you a barbarian. How do you answer them?**
> I am their king, not their teacher.

> **What advice do you give your son Alexander about ruling?**
> Be kind and gentle, but be careful not to let them know it.

## Running it

Q5_K_M GGUF, served with Ollama. Pair it with the elicitation frame (a "questioner puts
this to Philip" template); see the `Modelfile` in the project repo. Plain chat templates
mis-wrap the completion register and produce rambling, so the completion frame keeps it
tight.

## Intended uses

- The Elect console (Ollama serve plus chat)
- Historical dialogue and character embodiment
- Research into ancient Macedonian rhetoric and diplomacy

## Limitations

This is a plausible reconstruction, not a transcript. Philip's surviving direct speech is
thin (~28 apophthegms plus 23 letter sections), and the T2 layer extrapolates from
secondary sources. Because the corpus is so thin, the model can occasionally break register,
trailing into analysis or third-person commentary on the shortest prompts. Treat every
output as a plausible Philip, never as a historical quotation.

## License

CC-BY-NC-4.0. The model weights (GGUF) are released publicly under this license. The
training corpus is not bundled: the T1 sources are public domain and rebuildable from the
citations above (Perseus and Loeb), and the constructed T2 layer is generated text that is
not redistributed. Translations used (Babbitt 1931, Vince 1926) are US public domain.