Mapika/decider-2b-GGUF

decider-2b GGUF

GGUF files of Mapika/decider-2b v12, for llama.cpp. decider-2b does not generate text. It reads a state and one or more questions, each with an explicit option list, and returns a probability for every option from one forward pass. See the decider-2b card for what the model is, how it was trained, and where it is weak.

This repository holds the v12 files (2026-10-20). The v11 files stay available under the Hub tag v11. The GGUF files hold the text model only: image input needs the PyTorch weights of Mapika/decider-2b and decider-ai 1.10.0. decider_config.json reads yes/no (Noul) answers at temperature 0.05, as the model's own config does (see its card); decide_gguf.py reads every question as Choice, at the Choice temperature.

file	size	use
decider-2b-v12-Q4_K_M.gguf	1.3 GB	smallest; 0.3 points lower in-task and 0.4 held-out (table below)
decider-2b-v12-Q8_0.gguf	2.0 GB	same quality as the bf16 weights; the recommended file
decider-2b-v12-BF16.gguf	3.8 GB	unquantized, for making other quantizations

The tokenizer, decider_config.json (temperatures) and decide_gguf.py (the readout on llama.cpp) are in this repository too.

This is not a chat model

Loading a file in llama-cli, llama-server, Ollama or LM Studio gives you a text model that continues prompts. That is not how decider-2b is used, and its generated text is not its answer. The answer is read from the logits of the option-letter tokens at each answer slot of a prompt built by decider.prompt, divided by the fitted temperature. Decider (decider-ai 1.6.0) and decide_gguf.py do this with llama-cpp-python.

Usage

With decider-ai 1.6.0 or newer, Decider loads the GGUF file directly: decide, system_one and the per-type temperatures of decider_config.json, scored by llama.cpp.

pip install "decider-ai[gguf]"       # llama-cpp-python; GPU: CMAKE_ARGS="-DGGML_CUDA=on" (Apple Silicon: -DGGML_METAL=on)

from decider.infer import Decider

d = Decider("Mapika/decider-2b-GGUF", gguf_file="decider-2b-v12-Q8_0.gguf")
d.decide("My card was charged twice for the same purchase.",
         [{"question": "Which department should handle this?", "options": ["billing", "technical", "sales"]}])

gguf_options=dict(n_gpu_layers=0, n_threads=8) runs on the CPU. The standalone script below does the same decide() readout with decider-ai 1.5.0.

Standalone script (decide_gguf.py)

pip install decider-ai==1.5.0 llama-cpp-python      # llama-cpp-python 0.3.35 or newer
# GPU: CMAKE_ARGS="-DGGML_CUDA=on" pip install llama-cpp-python   (Apple Silicon: -DGGML_METAL=on)
hf download Mapika/decider-2b-GGUF --local-dir decider-2b-gguf \
  --include "*Q4_K_M.gguf" --include "*.json" --include "*.jinja" --include "*.py"
cd decider-2b-gguf

from decide_gguf import GGUFDecider

d = GGUFDecider("decider-2b-v12-Q4_K_M.gguf")      # n_gpu_layers=-1 (all on the GPU if the build has one), n_threads=...
d.decide("My card was charged twice for the same purchase.",
         [{"question": "Which department should handle this?", "options": ["billing", "technical", "sales"]},
          {"question": "How urgent is this?", "options": ["low", "medium", "high"]}])

decide_gguf.py covers decide() (choice questions); system_one with score and yes/no answers needs decider-ai 1.6.0 (above). The HTTP server does not serve GGUF files.

Measured quality

The regression set of the decider-2b card (95 tasks, 144,226 questions, 67 in-task and 28 held-out tasks) at the Choice temperature 2.465, read through llama.cpp (CUDA build, one prompt per decode) and compared row by row with the bf16 weights in PyTorch. Accuracy, NLL and ECE are means over tasks.

	in-task acc / NLL / ECE	held-out acc / NLL / ECE	same answer as bf16 PyTorch
bf16 weights, PyTorch	0.7906 / 0.5205 / 0.0532	0.7478 / 0.6440 / 0.0902
Q8_0	0.7912 / 0.5205 / 0.0530	0.7481 / 0.6440 / 0.0915	99.22%
Q4_K_M	0.7873 / 0.5304 / 0.0549	0.7442 / 0.6522 / 0.0907	95.08%

Q8_0 is equal to the bf16 weights within noise. Q4_K_M is 0.3 points lower in-task and 0.4 held-out. Use Q8_0 unless the smaller file matters. The BF16 GGUF was not read over this set. Long states (tens of thousands of tokens) move more under Q4_K_M than this table shows (measured on v11).

Notes

• Score one prompt per llama_decode, as decide_gguf.py does. With several prompts in one decode (as separate sequences), llama.cpp in September 2026 gives probabilities that change with the other prompts in the batch, by up to 0.02 in BF16 and 0.16 in Q4_K_M on the earlier version of this model. One prompt per decode gives the same numbers on every run.
• CPU and GPU builds give slightly different probabilities on the same file.
• Conversion: llama.cpp c9064dded (2026-09-27), convert_hf_to_gguf.py --no-mtp --outtype bf16, then llama-quantize to Q8_0 and Q4_K_M. --no-mtp is required: the checkpoint has no multi-token-prediction weights, but its config declares one MTP layer, and without the flag the converter writes a file that llama.cpp cannot load.
• The measurements above use the CUDA build. The CPU and Metal builds were not run over the regression set.

License: as decider-2b (see its card for the training sources and their terms); base model Qwen/Qwen3.5-2B-Base.