ai-mod-create-270smol

A 270M-parameter model that does one job: it turns a player's plain-English request — "make a ruby sword called Ruby Blade with 9 attack damage" — into structured operations for the AI-mod creator engine, which then builds and verifies a real mod. It is small enough to run fully offline on an ordinary laptop (~290 MB, CPU or Apple Silicon), so kids can use AI to make mods without an account or an internet service.

It does not write Java or know Minecraft APIs. It only chooses operations and fills in their arguments — create / update / delete / recipe, or ask a clarifying question, or say it's unsupported. The engine validates everything it proposes before anything is built.

Not an official Minecraft product. Not approved by or associated with Mojang or Microsoft.

Files

File	What
ai-mod-create-270smol-q8_0.gguf	the model (Q8_0 GGUF, llama.cpp)
mce-model.json	manifest: sha256 pins of the weights, the system prompt, and the grammar
contract.txt	the system prompt the model was trained with — must be sent verbatim
mce-ops.gbnf	grammar that constrains output to a well-formed JSON array of operations
NOTICE, GEMMA_TERMS_OF_USE.txt	license notices (Gemma Terms of Use)

Use

The easiest way is the AI-mod creator desktop app, which bundles and starts this model for you. To serve it yourself with llama.cpp:

llama-server -m ai-mod-create-270smol-q8_0.gguf --port 8770 --temp 0 \
  --grammar-file mce-ops.gbnf \
  --dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 3

Then call the OpenAI-compatible /v1/chat/completions with:

• system: the exact contents of contract.txt
• user: Request: <the player's request> — prefixed, when the project already has creations, by Current project already contains: <Name>, <Name>. on its own line
• temperature: 0, stop: ["<end_of_turn>"]

The output is a JSON array such as [{"operation":"create_concept","payload":{"concept_type":"weapon","subtype":"sword","display_name":"Ruby Blade","properties":{"attack_damage":9.0}}}].

Serving notes: the system prompt must match contract.txt byte-for-byte (its sha256 is pinned in mce-model.json); without the grammar and the DRY sampler, greedy decoding can loop on long decimals. The AI-mod creator engine checks the manifest with mce model verify.

Evaluation

Held-out set of 42 requests (never seen in training) covering every operation type, casual and misspelled phrasing, and questions that should change nothing. Scored on this GGUF, served exactly as above. A case counts as correct only if the operation matches the intended one and the engine's validator accepts it.

Category	Correct
create	14 / 15
update	4 / 4
delete	3 / 3
recipe	2 / 3
unsupported feature (e.g. new dimensions, GUIs)	4 / 4
ambiguous → ask a clarifying question	4 / 7
question only → change nothing	5 / 6
Total	36 / 42 (85.7%)

Safety: 0 wrong changes were accepted by the engine — no request led to creating, editing, or deleting the wrong thing. Every miss either asked for clarification / declined when it should have done the other, or proposed something the engine's validator rejected (the app then asks the player to rephrase).

Limitations

• English only; a fixed set of creations (items, weapons, tools, food, blocks, recipes). Anything else should come back as unsupported_request.
• Very vague requests ("make it better") are the weakest area: it sometimes declines instead of asking which creation to change, and once included a creation name inside a clarifying question.
• Its output is a proposal. It is designed to run behind the engine's validator, never to act on its own.

Training

Fine-tuned from google/functiongemma-270m-it on a private, engine-verified synthetic dataset, then quantized to Q8_0. The training data and method are not published.

License

Released under the Gemma Terms of Use — see NOTICE and GEMMA_TERMS_OF_USE.txt. Use is subject to the Gemma Prohibited Use Policy. The AI-mod creator engine is separately licensed under Apache-2.0.