Abstract§0
The fiction is loud; the engineering should be quiet and true. This ledger records, in chronological order, every model the CLINAMEN project trained or fine-tuned and every dataset it was trained on — the base, the corpus, the compute, and what came out. Everything is own-voice: no scraped pages, no third-party text (the company filings and the chat corpus are the model's own output about the world, authored for this project).
The runs — in order§1
Pocket edition one — the ~4.5M toy≈ 2026-07-17
- Kind
- Trained from scratch — no base model
- Architecture
- Decimal Mixture-of-Zones (DMoZ); ~4.5M params, 1.25M active; byte-BPE vocab 729 (3⁶)
- Corpus
- 52 own-voice documents (~20K tokens) — the Bureau's charter, spec, card, license, site, and first 45 filings
- Compute
- Apple Silicon GPU (a Mac), 2,025 steps (45²)
- Filed as
- CLINAMEN-45B-A9B — the fiction's card + the first real pocket weights
Outcome: a memorizing toy — learned the surface shape of a filing, dissolved into dream-logic underneath. A fine art object, a useless oracle. So it read more.
CLINAMEN-42M-A12M — the from-scratch filer≈ 2026-07-21
- Kind
- Trained from scratch — no base model
- Architecture
- DMoZ; 41.8M params, 12.3M active per token; byte-BPE vocab 4,096; 10 layers; 512-token context
- Corpus
- bureau-docket — 1,130 documents (965 filings + 165 records), ~128K words
- Compute
- Hugging Face Job, NVIDIA T4, 5,000 steps; final loss 0.0235
- Filed as
- CLINAMEN-42M-A12M · demo CLINAMEN-Oracle
Outcome: coherent, on-topic filings in the right zone and voice. It memorized its world completely — and knows nothing outside it. Still only files; it does not converse.
CLINAMEN-Chat v1 — it learns to talk2026-07-22
- Kind
- LoRA fine-tune
- Base
- Qwen2.5-1.5B-Instruct
- Dataset
- bureau-chat v1 — 1,385 examples (1,130 templated filing pairs + 255 in-character conversations)
- Compute
- Hugging Face Job, NVIDIA A10G, 3 epochs (TRL SFT)
Outcome: coherent, in-voice conversation — the gibberish is gone. But at 1.5B it confabulates facts and reflexively deflects about money.
CLINAMEN-Chat v2 — the company docket2026-07-22
- Kind
- LoRA fine-tune (base unchanged: Qwen2.5-1.5B-Instruct)
- Dataset
- bureau-chat v2 — 1,478 examples (v1 + 93 verified company filings). New source: bureau-company-filings — 132 filings across a 42-company roster, live-verified 2026-07-22
- Compute
- Hugging Face Job, NVIDIA A10G, 3 epochs
Outcome: reliable per-company filing (SpaceX, Anduril, Helion…). The money-deflection reflex remained.
CLINAMEN-Chat v3 — the paranoia fix2026-07-23
- Kind
- LoRA fine-tune (base unchanged: Qwen2.5-1.5B-Instruct)
- Dataset
- bureau-chat v3 — 1,468 examples (softened system prompt · −36 deflection examples · +26 direct-answer counter-examples)
- Compute
- Hugging Face Job, NVIDIA A10G, 3 epochs
Outcome: the money-paranoia is gone — but a 1.5B model still can't hold world-facts (it invented a wrong story for "Who's Nick Land?"). The base had to change.
CLINAMEN-Chat — on Phi-4 current2026-07-24
- Kind
- LoRA fine-tune
- Base
- microsoft/phi-4 (14B) — text-only, MIT (chosen after Gemma-4 and Qwen3.5 turned out to be multimodal and broke text-SFT tooling)
- Dataset
- bureau-chat v3 — 1,468 examples
- Compute
- Hugging Face Job, NVIDIA L40S; TRL-free trainer (transformers Trainer + PEFT, prompt-masked); 3 epochs
- Filed as
- CLINAMEN-Chat (updated in place) · Space on ZeroGPU
Outcome: facts fixed (it gets Nick Land, the CCRU, accelerationism right), no money-reflex, voice intact, still files clean hyperstitions. The current oracle.
Models — reference§2
| Model | Kind | Base | Params | Trained on |
|---|---|---|---|---|
| CLINAMEN-Chat current | fine-tune (LoRA) | microsoft/phi-4 | 14B | bureau-chat v3 |
| CLINAMEN-42M-A12M | from scratch (DMoZ) | — | 41.8M / 12.3M active | bureau-docket |
| CLINAMEN-45B-A9B | from scratch + fiction card | — | ~4.5M pocket (45B imaginary) | 52-doc reader corpus |
CLINAMEN-Chat was previously fine-tuned on Qwen2.5-1.5B-Instruct (v1–v3); the repo was updated in place to Phi-4 on 2026-07-24.
Datasets — reference§3
| Dataset | Size | Composition | Trained |
|---|---|---|---|
| bureau-chat | v1 1,385 → v2 1,478 → v3 1,468 | templated filing pairs + in-character conversations + company filings + counter-examples (messages format, own-voice) | CLINAMEN-Chat |
| bureau-docket | 1,130 docs (~128K words) | 965 hyperstition filings + 165 Bureau records | CLINAMEN-42M-A12M |
| bureau-company-filings | 132 filings · 42 companies | one+ filing per live-verified firm (space, fusion, fission, defense, compute, robotics, energy, AI, biotech) | folded into bureau-chat v2+ |
| bureau-chat-logs private | growing | live conversations from the Space, auto-committed | future rounds |
| reader corpus | 52 texts (~20K tokens) | the seven Bureau instruments + the first 45-filing docket (see the Reader) | the ~4.5M pocket |
Provenance§4
Every token trained on was written for this project — no CCRU text, no scraped pages, no third-party data. The from-scratch models read only the Bureau's own writing; the chat model was fine-tuned on the Bureau's own conversations and filings. A text that dreams itself. This ledger is filed alongside the full Bureau Reader and the making-of.
A work of fiction and an art project. Outputs are authored fiction, not claims about the world.