What it was trained on
118 rows from 101 sketches the base model wrote that Adam rated "love" (17) or "okay" (84),
out of the ones he judged blind. Each love appears twice. Each row is the system prompt in
system_prompt.txt, a user request, and the sketch the base model wrote for it.
This is filtered behaviour cloning: the model learns from its own outputs that a person liked.
Training
Table with columns: Setting, Value| Setting | Value |
|---|
| Base | Qwen/Qwen3.5-35B-A3B (language model only; vision modules excluded) |
| LoRA | rank 16, alpha 32, dropout 0.05, on q_proj, k_proj, v_proj, o_proj |
| Optimiser | learning rate 1e-4, cosine decay, 5% warm-up, gradient accumulation 4 |
| Length | 3 epochs, 88 steps; mean loss by epoch 0.175, 0.163, 0.155 |
| Hardware | one AMD Strix Halo box (gfx1151, ROCm) |
Files
adapter_model.safetensors, adapter_config.json: the PEFT adapter.
adapter.gguf: the same adapter converted for llama.cpp. This is the file the paintings were made with.
system_prompt.txt: the system prompt it was trained and run under (called M2 in the project).
How to run it
The paintings were made with llama.cpp, a Q8_0 GGUF of the base model, thinking off, temperature 1.0:
llama-server -m Qwen3.5-35B-A3B-Q8_0.gguf --lora adapter.gguf
Send system_prompt.txt as the system message and a request like this as the user message:
Write a p5.js sketch that paints an original Mondrian-style grid composition with a tall red planeon the left edge and a small yellow plane lower right. Use only the allowed functions. Do notreproduce any existing painting.
The reply is a p5.js sketch on an 800 × 800 canvas. Paste it into the
p5.js editor to see the painting.