npateman/fusewalk-models

Fusewalk models

Fine-tuned models for Fusewalk, a crafting game: give a model two items and it answers what they make.

Each model is a 4-bit (Q4_K_M) GGUF for llama.cpp. index.json lists them with the exact prompt template and stop strings each one needs.

Modifications

Each model is a fine-tune of the base model listed below. It was trained (LoRA merged into the weights, or a full fine-tune) on the Fusewalk Codex, a hand-written set of item-combination recipes, so that it answers with the names of the items two items make. It was then converted to GGUF and quantised to 4-bit. The base models' licences and notices are in licenses/, copied unchanged from the base model repos.

Models

Model	Fine-tune of	Training	Licence	Codex	Size	Plausible
20261016-152956-qwen3-1.7b-v0.6	Qwen/Qwen3-1.7B	LoRA	apache-2.0	v0.6	1,056 MB	51%
20261015-090302-qwen3-1.7b-v0.4	Qwen/Qwen3-1.7B	LoRA	apache-2.0	v0.4	1,056 MB	57%
20261015-090921-qwen3-0.6b-v0.4	Qwen/Qwen3-0.6B	full fine-tune	apache-2.0	v0.4	378 MB	41%
20261014-135949-qwen3.5-4b-v0.4	Qwen/Qwen3.5-4B	LoRA	apache-2.0	v0.4	2,583 MB	74%

Plausible is the share of answers that make sense, on pairs held out from training: an LLM judge rates each answer good, mash-up (the two names glued together), weak or nonsense, and this counts the good ones. There's often more than one sensible answer, so it says more than an exact match with the Codex would.

Try it

The prompt is two quoted item names, in alphabetical order, joined with +, inside the model's prompt template. The answer is one or two quoted names on the first line. With llama.cpp and the newest model:

llama-completion -m 20261016-152956-qwen3-1.7b-v0.6.gguf -no-cnv --temp 0 -n 24 \
  -p '<|im_start|>user\n"Fire" + "Water"<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n'

It answers something like:

"Steam"

Each model's template is prompt_template in index.json, with {pair} where the items go. Stop at the first newline.