Modelpublic

SebastianAldrin/Qwen3.5-9B-minecraft-distill-v1-GGUF

sourceHugging Facemitupdated 2mo agoView on Hugging Face

2likes64downloads

Model Card

Qwen3.5-9B — Minecraft Agent Distill v1 (GGUF)

A LoRA fine-tune of Qwen/Qwen3.5-9B that distils Claude Sonnet's per-tick decisions in the Agent Society Minecraft sandbox — so the agent's fast decision tier runs locally and offline instead of calling a frontier model every tick.


Base	`Qwen/Qwen3.5-9B`
Method	LoRA SFT — rank 16, alpha 32, dropout 0.05, all-linear, 3 epochs
Teacher	Claude Sonnet
Data	SebastianAldrin/agent-society-distill-v1 — 1,355 examples
Code	Agent Society
Format	GGUF, Q4_K_M (~5.4 GB) — runs in llama.cpp / Ollama

Given the agent's situation as a prompt (felt needs, a local block-map, bearings, the current plan step, recent memory, who else is nearby), it returns one in-character decision as JSON: {"thought": "...", "action": "...", "args": {...}}. It is the fast per-tick tier of a three-tier agent mind; planning and reflection stay on a stronger model.

Run

bash

llama-server -m Qwen3.5-9B-minecraft-distill-v1-Q4_K_M.gguf -c 8192 --jinja

Send the decide prompt with enable_thinking: false and a JSON-schema response format (the model is trained to answer with the decision JSON only). Full system prompt and prompt format live in the Agent Society repository.

On CUDA, add --flash-attn off — llama.cpp's auto default silently corrupts this hybrid-SSM architecture: the JSON shape survives but the words inside turn to noise.

Limits

—Trained on a small, mostly-social set (1,355 examples), so it talks a lot and is weak at rare actions.
—Needs a recent llama.cpp that knows the qwen3_5 arch — older builds won't load it.
—The prompt isn't cacheable on this arch, so it re-reads the whole prompt every tick. Slow on a weak GPU or CPU.
—Measured: on 250 held-out teacher decisions it picks the teacher's action 78.0% of the time; the untuned base scores 56.8%. Method in the evaluation doc.

License

MIT.