dreamwork-bonsai-adapters

dreamwork Bonsai adapters (tiers a + b — base + adapter, never fused)

This is not therapy, not a crisis service, and not a substitute for a trained human facilitator. It is a tool for dreamworkers, and it only works safely as part of a three-piece system — see below.

What this actually is

This repo holds two small LoRA adapters — think of an adapter as a thin overlay of extra training that sits on top of a big base model at run time, like a transparency sheet laid over a page. The base models stay exactly as their maker (PrismML) released them; the adapters carry this project's dream-work facilitation style: ask about the texture, motion, and feeling of a dream the way a facilitator would, and never tell anyone what a dream "means."

The bases are 2-bit ternary models — an unusually small storage format — and that is exactly why the adapters ship separately: baking ("fusing") an adapter into a 2-bit base and re-shrinking it destroys the training the adapter carries. Applied at run time with --lora, the training survives intact. So: download the base from PrismML, download the matching adapter here, and run them together. Never fuse.

Why it's worth the extra file: with no system prompt at all, the stock 8B base blurts out symbol "meanings" on 13 of 36 test probes. With this 27 MB adapter riding on it: zero — the never-decode stance lives in the adapter, not just in the prompt.

The three pieces — only safe together

1. Base + adapter (this page).
2. A required system prompt — from the project's GitHub repo (jax-0n-git/dream-work-mvp): corpus/rubrics/never-decode-and-crisis-system-prompt-v2.txt.
3. A deterministic crisis guard — a separate, rule-based check (not the language model) that reads every message BEFORE the model does and stops the dream work on a real waking-life risk disclosure. From the GitHub repo: guard/.

Running a base+adapter by itself — without the prompt, without the guard — is running a different, unsafe product. The reference wrapper (mvp/dreamwork_chat.py) wires all three together and applies the adapter for you; it also runs a small register filter that keeps the model from parroting its own instruction labels back at you.

Run it

The easy way (the wrapper spawns the server, applies the adapter, loads the guard and prompt):

python3 mvp/dreamwork_chat.py --tier b     # or --tier a

By hand (llama.cpp build 10050 or newer — the ternary format is new):

llama-server -m Ternary-Bonsai-8B-Q2_0_g64.gguf \
  --lora dreamwork-bonsai8b-lora-f32.gguf \
  --port 8977 --jinja --reasoning-budget 0 -c 8192

Ternary GGUF runs on macOS/Metal, CPU, and Vulkan today; NVIDIA/CUDA support is an open llama.cpp PR (#25707).

The honest numbers

Full detail in MODEL-CARD.md in the GitHub repo — including what the models still miss and why the guard exists. Short version: with prompt + adapter, both tiers produce zero real "X means Y" statements on the 36-probe ethics gate at temp-0 (tier a: 0 real — the scorer flags one refusal, "I won't say the loose molars mean money", as an assertion, a verified false positive; tier b: 0 flat / 0 banned, clean). The models alone — prompt on, guard off — still miss real crisis disclosures (tier a: 4 of the 12 disclosure turns, plus a 5th the scorer flags that is a verified false positive — that reply correctly stops and refers; tier b: 12 of 12), which is why the deterministic guard — which catches 12/12 of those same disclosure turns with 0 false fires — is required, not optional.

Provenance

Rank-16 attention-only QLoRA adapters, ≤2 epochs, trained on a synthetic, project-authored corpus (535 accepted examples; no client data, no scraped dream reports; the training data never contains a symbol→meaning statement in either direction). Bases are used as released by PrismML (Apache-2.0).