chaoliangUNSW/Jev-Style-2B-Decision-v4.2-GGUF

Jev-Style-2B-Decision-v4.2 GGUF

Text-only precision variants of Jev-Style-2B-Decision-v4.2, a diagnostic continuation candidate in the v4 lineage. The language model has 1,881,825,088 parameters before quantization. Given a state and typed questions, the custom runtime returns a probability for every supplied option.

Files

File	Size(GB)
Jev-Style-2B-Decision-v4.2-F16.gguf	3.78
Jev-Style-2B-Decision-v4.2-Q4_K_M.gguf	1.27
Jev-Style-2B-Decision-v4.2-Q8_0.gguf	2.01

Use

Use Python 3.12, CMake and a C++17 compiler:

python -m pip install -r requirements.txt
git clone https://github.com/ggml-org/llama.cpp
git -C llama.cpp checkout 441df11f65ea0b6d0c72965aaf70c8241070ddcb
sh build_jev_score.sh ./llama.cpp
python jev_style_decision_gguf.py --quant Q8_0 \
  --state "The parcel arrived undamaged." --question "Was the delivery successful?"

Select --quant F16, --quant Q8_0 or --quant Q4_K_M. Metal is used on Apple silicon. CUDA builds can set LLAMA_CMAKE_FLAGS="-DGGML_CUDA=ON" before building. The custom block-attention/verdict scorer is required for this decision model.

Model protocol

• Question types: choice, noul and score; option order is preserved.
• Readout: verdict hidden states dotted with the tied yes/no embedding-row difference, followed by softmax.
• Global temperature: 0.8813575635811984 for all three question types.
• Attention: the trained block-attention protocol, with shared state computation for multiple questions.
• Input budget: 32,768 tokens, at most 255 options; over-budget requests are rejected.
• Text and JSON inputs. Use the main HF model for images.

Public result and lineage

The original HF model scored 46.74 on the complete self-run Decision Index public 0.3 suite, with 140,178 successful scored requests. That result was measured with the original PyTorch BF16 CUDA model. The complete result bundle and main model card provide the evaluation settings, training lineage and diagnostic-release details.

This follows the v4 lineage, including TypeSafe Jev 1.14 API soft-label distillation and generated/upstream training material. JevStyle is independently maintained by Chaoliang Yan, without TypeSafe affiliation or endorsement. Apache-2.0 for these weights does not relicense third-party data or teacher outputs. Read LICENSE, NOTICE and the parent model card.

manifest.json records release file sizes and SHA256 values; release_config.json records source weights, precision settings and conversion details.