JJ-1-2B

日本語 · ← Back to JJ-1 homepage · JJ-1 article: English / 日本語

Japanese business decisions from text or JSON: choices, yes/no probabilities and scores. Includes BF16, Q8_0 and inference code.

Read the full JJ-1 technical report in English or Japanese for the evaluation setup and detailed results. Register for the JJ-1 inference service.

Evaluation

BF16 results from the JJ-1 report (88,813 questions, 6 packages):

Evaluation	Hard-label accuracy
Overall	68.54%
English	78.09%
Japanese	62.63%
Business workflows · test + held-out	94.63%

Business results combine test and held-out cases. Q8_0 has not been independently benchmarked.

Six-area comparison

Comparison of the 17 models with complete results across all six areas in the report. Values are hard-label accuracy (%); Overall is weighted by question count. Soft-label questions are excluded from this table.

Column	Evaluation area	All questions	Hard-label questions
01	Japanese enterprise workflows	6,249	6,033
02	News, tools & safety	10,967	10,967
03	Japanese knowledge & reasoning	45,391	45,391
04	English knowledge & classification	22,417	19,115
05	Emotion & scenario classification	2,200	2,200
06	Japanese semantic similarity	1,589	447

Model	Overall	01	02	03	04	05	06
JJ-1-9B	79.20%	97.43%	92.17%	72.70%	81.31%	80.50%	79.19%
StartLux 9B	78.02%	90.10%	92.36%	71.29%	81.73%	80.86%	73.60%
Jev 1.13	76.50%	91.75%	89.81%	71.96%	76.82%	60.32%	70.92%
Microsoft Decision-1	76.50%	91.58%	90.53%	69.76%	81.22%	63.45%	74.50%
JJ-1-4B	75.18%	96.17%	90.89%	67.29%	78.72%	70.41%	78.97%
StartLux 4B	74.54%	89.34%	91.47%	66.55%	79.63%	71.68%	67.34%
JPT 9B	73.45%	89.36%	90.43%	66.51%	76.54%	61.86%	71.14%
Nimble 9B v3	72.00%	89.21%	84.95%	65.04%	77.08%	60.32%	69.13%
JevK5 9B	71.77%	87.58%	89.09%	65.16%	74.15%	58.27%	69.35%
GPT-6 Luna Decisions	71.35%	91.99%	86.79%	65.07%	72.16%	59.41%	75.17%
Jebadiah 9B v2	70.82%	88.36%	87.72%	63.09%	75.49%	59.50%	61.74%
JPT 4B	70.62%	87.15%	90.26%	62.32%	75.15%	61.86%	57.72%
Kev 9B	70.08%	85.61%	89.24%	62.37%	74.13%	57.27%	63.98%
JJ-1-2B	68.54%	94.63%	87.35%	58.19%	73.42%	73.00%	75.39%
StartLux 2B	68.24%	82.51%	89.22%	57.78%	75.80%	75.55%	63.98%
StartLux 0.8B	65.64%	76.99%	88.91%	53.74%	74.44%	88.23%	63.31%
JJ-1-0.8B	65.49%	92.31%	87.33%	52.43%	72.62%	88.50%	75.62%

Column 06 is the report’s hard-label accuracy, not a semantic-similarity correlation coefficient. These are report results, not an independent Q8_0 benchmark.

Transformers

Python 3.10+:

cd JJ-1-2B
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[transformers]"
python example.py

from jj1 import JJ1

model = JJ1(".", images=False, device="cpu", max_length=4096)
state = "洗面所の天井から水が落ち続けています。"
questions = {
    "needs_visit": {
        "type": "noul",
        "instructions": "現地での対応が必要ですか？"
    }
}
answers, usage = model.decide(state, questions)
print(answers)
print(usage)

CLI / HTTP

From the model folder:

jj1-check .
jj1-decide --model . --device cpu --pretty examples/request.json
jj1-serve --model . --device cpu --no-images --max-length 4096 --port 8090

HTTP: POST /v1/systemone.

GGUF / llama.cpp

In the activated environment:

python -m pip install -e ".[gguf]"
llama-server -m JJ-1-2B-Q8_0.gguf -c 4096 --parallel 1 \
  -ngl 0 --host 127.0.0.1 --port 8081 --no-webui

Second activated terminal, same folder:

python example.py --backend gguf

If 8081 is occupied, change --port and use example.py --url http://127.0.0.1:PORT.

GGUF requires llama-server; PyTorch and Transformers are unnecessary.

Build llama-server

Build outside the model folder with Git, CMake and a C++ compiler:

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git checkout b7dafa01e5f375c3010fb24b61a67329f957426a
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_OPENSSL=OFF
cmake --build build --config Release -j 4 --target llama-server
export PATH="$PWD/build/bin:$PATH"

NVIDIA CUDA

Requires CUDA PyTorch and the fast kernels:

python -m pip install -e ".[transformers,cuda]"
python example.py --device cuda

Notes and license

Text/JSON only.

Weights: CC BY-NC 4.0 (LICENSE). Code: Apache-2.0 (LICENSE-CODE). Attribution: NOTICE.