Modelpublic

lewismoten/palace-9

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face

0likes390downloads

Model Card

Palace-9 raw-history model

A compact, from-scratch Qwen2MoeForCausalLM model for a 3×3 tic-tac-toe move-history task. It is a raw, one-token state-completion model—not a general conversational model.

Live browser demonstration: lewismoten.github.io/palace-9

Source repository: lewismoten/palace-9

Browser presentation label: PALACE — Predictive Autonomous Learning And ~~Nuclear~~ Contingency Evaluation. The struck word is intentional fictional presentation; the release is a 3×3 game-state model with no operational capability.

[image]

Browser weight inspector. The interactive site decodes the selected Palace-9 release artifact locally and visualizes its causal-model forward path. The playable board, map, DEFCON display, trajectories, and cipher are browser presentation layers—not Hugging Face inference or an operational system.

Model contract

Submit only a raw move history made from a through i—one letter for each occupied square, in chronological order. The deployment template is literal <bos>{{ .Prompt }}. Use temperature 0, a 16-token context, and generate exactly one token.

—A legal history returns an optimal unoccupied square.
—A malformed, repeated-square, post-terminal, or otherwise out-of-protocol history returns !.
—This model does not implement a multi-turn chat protocol.

Example raw history: a Checked F16 fixture result: e

Architecture

—Architecture: Qwen2MoeForCausalLM, trained from scratch
—Vocabulary: custom 261-token byte-level GPT-2-compatible vocabulary
—Context: 16 tokens
—Decoder layers: 1
—Hidden size: 36
—Attention: 9 query heads × 4 dimensions; 3 KV heads × 4 dimensions
—Routed experts: 9; top-2 routing
—Shared expert: 1

Release artifacts

Artifact	Runtime evidence
F16 GGUF	978,003 raw-history cases; 0 failures; local Ollama fixture verified
Q6_K GGUF	978,003 raw-history cases; 0 failures
Q4KM GGUF	978,003 raw-history cases; 0 failures
Source checkpoint	Hugging Face-compatible config, tokenizer, and `safetensors` weights

The Q4KM release is truthfully mixed storage: narrow tensors that cannot use a particular block layout remain F16/F32 or Q6K where required. No Q80 artifact is provided because the model's 36- and 18-wide tensors do not meet Q8_0 block-size requirements.

Use with Ollama

Use the F16 GGUF with a raw-completion template:

text

FROM ./palace9-qwen2moe-raw-history-f16.gguf
PARAMETER num_ctx 16
PARAMETER num_predict 1
PARAMETER temperature 0
TEMPLATE """<bos>{{ .Prompt }}"""

Then import with ollama create palace-9:f16 -f Modelfile.

The Ollama package contains only the GGUF and Modelfile configuration. The playable board, weight inspector, and fictional visual overlay run independently in browser JavaScript and are available through the source repository and live demo.

Safety and scope

This is a small, fictional 3×3 tic-tac-toe state policy. It has no external command authority, no real-world data, and no operational-system capability. The browser demo's map, DEFCON display, trajectories, and cipher graphics are fictional presentation layers.

Provenance and validation

The release includes source checkpoint files, GGUF artifacts, checksums, a model card, and machine-readable validation evidence. The exhaustive runtime gates cover 294,777 legal histories and 683,226 invalid histories (978,003 total) per F16, Q6K, and Q4K_M artifact, with zero failures.

Acknowledgments

PALACE-9 was designed and directed by Lewis Moten. Its code and documentation were developed with assistance from GPT-5.6-terra Med, accessed through Hermes and using Honcho for context and project-memory support. Lewis Moten remains the project designer, maintainer, and publisher.