χρυσοποιία — the alchemical making of gold. An exploration project: soak a small, fully-open base model in public-domain Western-esoteric prose until the esoteric register becomes its default completion dialect, then install turn-taking with a light SFT — so it speaks in a spiritual/esoteric voice by default, on any input, with no retrieval at inference.

This is not a general assistant and not a knowledge oracle. It is a study of whether a disposition (and some conceptual substance) can be baked into weights. It will state esoteric "facts" fluently whether or not they are accurate — do not treat it as authoritative.

📖 How it was built (the full story — what worked, what didn't, and the one idea it turned on): https://github.com/4-R-C-4-N-4/chrysopoeia/blob/master/docs/how-it-was-built.md

What it does

Ask it anything — mundane or esoteric — and it answers in the register of the early-20th-century Western esoteric revival (Manly P. Hall, Éliphas Lévi, the Kybalion), while still giving a real answer:

Q: My car won't start this morning. What should I check? The vessel that carries you into the day has lost its fire — yet even the cold stone remembers its own hidden heat… Begin at the obvious fount: the battery's charge, the terminals' grip, the fuel's supply. If these pass their test, seek the hidden: a coil grown old, a circuit severed by time…

Method (two phases)

Soak — QLoRA continued-pretraining of SmolLM3-3B-Base on ~2.6M tokens of public-domain Western-esoteric prose (constant LR; a mid snapshot is used, not the final one, which memorizes). This installs esoteric substance.
Light SFT — a small turn-taking pass on a mundane-input → in-register response slice. This is what severs topic from register, making the voice the default rather than a topic-triggered reflex. Soak alone does not do this; the mundane slice is the load-bearing move.

Prompt format

A minimal plain-text chat format (no special tokens):

### User:
{your question}

### Chrysopoeia:

Generation should stop at the next ### User:.

Run it (llama.cpp)

Download a GGUF (gguf/chrysopoeia-smollm3-Q4_K_M.gguf, ~1.9 GB) and serve it:

llama-server -m chrysopoeia-smollm3-Q4_K_M.gguf -c 2048 -ngl 999

Then hit the raw /completion endpoint with the plain-text format (it has no chat template — use /completion, not /v1/chat/completions):

bash

curl -s http://127.0.0.1:8080/completion -d '{
  "prompt": "### User:\nMy car won'\''t start this morning. What should I check?\n\n### Chrysopoeia:\n",
  "n_predict": 200, "temperature": 0.7, "top_p": 0.9,
  "stop": ["### User:"]
}' | python3 -c "import sys,json;print(json.load(sys.stdin)['content'])"

GGUF options

File	Size	Use
`gguf/…-Q8_0.gguf`	3.3 GB	near-lossless fidelity
`gguf/…-Q4_K_M.gguf`	1.9 GB	default — no meaningful loss
`gguf/…-IQ3_M.gguf`	1.47 GB	small; imatrix-guided, voice intact (+7% ppl)
`gguf/…-f16.gguf`	6.2 GB	full precision

chrysopoeia-esoteric.imatrix is the importance matrix (calibrated on the model's own esoteric distribution) — use it to roll your own IQ quants. See the build repo's docs/quantization.md for the size-vs-quality numbers.

Also: merged/ (bf16 safetensors for 🤗 Transformers), adapters/ (the composing soak + SFT LoRAs). Ollama: ollama run 4rc4n4/chrysopoeia.

Training data & provenance

Soak corpus: confirmed US public-domain Western-esoteric prose only (Hall's Secret Teachings of All Ages 1928, the Kybalion, Lévi's Transcendental Magic, Waite, Papus, Ouspensky; plus PD Hermetica / Neoplatonism / Gnosticism). Source status is load-bearing because the soak bakes prose into weights.
Mundane slice: ordinary human-written prompts (Dolly-15k, CC-BY-SA-3.0), answered in-register by a local model grounded in real corpus passages (retrieval-augmented generation at training time). Prompts filtered off the esoteric manifold so the register generalises as disposition, not topic-match.

Limitations

Voice over correctness. Grading targets the quality of the voice, not doctrinal accuracy. It will confidently fabricate esoterica.
Not a general assistant. General capability is expendable by design.
Experimental. Register depth, coherence, and the substance/coherence knee are all under active study; this is a checkpoint, not a finished artifact.

License

Weights derive from SmolLM3-3B-Base (Apache-2.0). Soak sources are US public domain; the mundane-slice prompts are Dolly-15k (CC-BY-SA-3.0). Released under Apache-2.0.

4rc4n4/chrysopoeia-smollm3

Chrysopoeia — SmolLM3-3B, soaked into the esoteric register