ai-wasteland-local

ai-wasteland local model

This is the compact local model used by ai-wasteland, a Minecraft Forge mod whose terrain, settlements, and monuments are authored as validated tool calls. It is a narrow structured generator, not a chat model or a general prose model.

The release contains a 52.3 MB Burn MessagePack model with 26,164,741 parameters. The 6-layer, 512-wide transformer uses grouped-query attention with 8 query heads and 2 key/value heads. Its 8,197-token vocabulary includes compact JSON macros plus byte fallback for previously unseen UTF-8 prompt text; those macros decode to the unchanged tool schema. Schema rules are shipped beside the weights. Saved generation defaults use temperature 0.25, top-k 2, a 1,400-token ceiling, and schema-aware JSON constraints.

Measured behavior

In the saved 12-prompt production-shaped evaluation, the selected continuation produced 12/12 parseable constrained and unconstrained samples. The real validator accepted all 12 in each mode, every sample answered the requested kind, flavour, band and purpose, and all ten monument/settlement samples followed their requested attachment plan. These figures are a small deterministic engineering check against the checkpoint's recorded corpus, not a general benchmark.

Use harness/verify-model.sh in the source checkout to reproduce the offline comparison. harness/verify-local-model.sh is the stronger release gate: it starts the model endpoint and a fresh headless Minecraft server, then requires the real sidecar to validate and publish the target pool. The release candidate passed that gate after source commit ed948aa: the fresh world published 3 terrain bases, 3 settlements and 10 monuments, completed two exact requested terrain/settlement pairs, one assistant turn executed 6 tool calls, and every accepted template scored at least 77/100. Its maximum sampled four-block terrain step was 8 against a limit of 12, and no terrain used architectural fittings.

Run it

Clone the source at the commit in SOURCE_COMMIT, download the files under model/ into trainer/data/model, then run on an NVIDIA host:

export AIW_API_TOKEN="$(openssl rand -hex 32)"
AIW_MODEL_DIR=trainer/data/model AIW_INFER_BACKEND=cuda ./model.sh serve 18080

The endpoint implements authenticated GET /v1/models and POST /v1/chat/completions. It returns one or more write_template tool calls and supports up to six independent decoder calls in a batch (the end-to-end run executed six tool calls from one assistant turn). The sidecar configures it through AIW_OPENAI_BASE_URL, AIW_OPENAI_API_KEY, and AIW_MODEL.

Production-shaped asks are in examples/prompts.jsonl. The serving adapter is part of the harness because these weights are not compatible with Transformers or llama.cpp.

Training and provenance

The selected checkpoint is the E10 one-epoch shrine-relation continuation of the matched E8 geometry candidate, with physical batch two, accumulation four, sequence length 2,368, sink-plus-window sparse attention (16 sink tokens and a 1,024-token window), and grammar unlikelihood weight 0.5. Its 3,815 balanced rows combine validator-proven topology corrections with the selected external building corpus. No generic prose rows or prose adapter were used. model/run.env records the realized settings and training source commit. Task queries include direct requests, multi-target pool status, validator retries, and critiques.

The corpus and raw experiment log are intentionally not in this model repo. Blueprint provenance and corpus construction live in the separate ai-wasteland-data source repository; validation and rating belong to the mod source so there is only one contract.

The real-building half of the corpus comes from three collections — GrabCraft blueprints, Meta's 3D-Craft houses, and the GDMC settlement generation competition — each validated and scored by this mod's own validator before anything is kept. How they are fetched, converted and licensed is recorded in ai-wasteland-data; 3D-Craft ships no licence of its own and should be treated as non-commercial research material.

Limitations

• Byte fallback accepts unseen UTF-8 prompt text, but the small model has not learned every concept or language represented by arbitrary input.
• Constrained decoding guarantees JSON/schema-local properties, not good buildings. Purpose, support, coherence, and cross-field requirements still need the production validator and retry loop.
• The checkpoint sometimes confuses template kinds or writes incoherent fields. Do not consume its output without the ai-wasteland validator.
• No standalone Transformers integration or hosted inference widget is provided.

See REFERENCES.md for source, data, and research references. SHA256SUMS covers every file in this release.