mitsutani/mahjonglm-10m-q4-k-m-gguf

MahjongLM 10M Q4_K_M GGUF

This repository contains the Q4_K_M GGUF export of mitsutani/mahjonglm-10m.

GGUF file:

• mahjonglm-10m-Q4_K_M.gguf

Training

The source model was trained on mitsutani/mahjonglm-dataset, using complete, imperfect, and omniscient MahjongLM views over Tenhou logs from 2011 through 2024.

Prompt Format

<bos> rule_player_4 rule_length_hanchan view_complete game_start

For view_omniscient, use:

<bos> rule_player_4 rule_length_hanchan view_omniscient game_start round_start wall

and then provide or generate the 136 wall tile tokens.

llama.cpp WordLevel Tokenizer Patch

MahjongLM uses a fixed WordLevel tokenizer over Mahjong log tokens rather than a byte-pair or sentencepiece tokenizer. The GGUF file contains the vocabulary metadata, but upstream llama.cpp builds may not know how to interpret this tokenizer type.

This repository includes the minimal llama.cpp patch in patches/:

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
git apply /path/to/patches/0001-wordlevel-tokenizer-support.patch
cmake -B build
cmake --build build --config Release

After building the patched binary, run the model normally with llama-cli or compatible llama.cpp tools. The patch is tokenizer-only; the model architecture is plain Qwen3.