basal-2.0-max-GGUF

GGUF files (Q8_0 and Q5_K_M) for Ollama and llama.cpp of basal-2.0-max (11B, base Bielik-PL-11B-v3.0-Instruct). Same model, same typed-decision API through basal-serve; the calibration files this folder ships are listed below.

Measured against bf16

Agreement is the share of decisions where the port picks the same option as the bf16 release model on the development set dev v2 (MLX rows: every 8th dev v2 decision, 1,700 decisions). Bars for this tier: agreement 2-10 >= 0.9552, agreement at reference margin >= 0.2 >= 0.9893, accuracy within 1 point of bf16, ECE at most +0.01. All ports of this tier:

The ports are decision-layer checks against bf16, not new benchmarks; the model's own results are on basal-2.0-max.

Use

Files: basal-2.0-max-Q8_0.gguf (recommended), basal-2.0-max-Q5_K_M.gguf (smaller), Modelfile.Q8_0, Modelfile.Q5_K_M, basal.json and the calibration files below. The basal server runs in front of Ollama or llama.cpp: it renders basal's own prompt and reads the log-probabilities of the option codes. Ollama 0.12.11 or newer:

uv venv --python 3.12 --allow-existing ~/basal-env && source ~/basal-env/bin/activate
uv pip install "basal @ https://github.com/rkinas/basal/archive/refs/tags/v2.5.3.tar.gz"
hf download Remek/basal-2.0-max-GGUF --local-dir basal-2.0-max-GGUF
cd basal-2.0-max-GGUF && ollama create basal-2.0-max:q8_0 -f Modelfile.Q8_0 && cd ..
basal-serve --model ./basal-2.0-max-GGUF --mode ollama --ollama-model basal-2.0-max:q8_0 --port 8000

--model points at the downloaded folder for the tokenizer, chat template and calibration files; Ollama serves the weights. Ollama returns at most 20 log-probabilities per position, so questions with more than 20 options need llama.cpp:

llama-server -m basal-2.0-max-GGUF/basal-2.0-max-Q8_0.gguf -c 4096 --port 8080
basal-serve --model ./basal-2.0-max-GGUF --mode llamacpp --llamacpp-url http://127.0.0.1:8080 --port 8000

Without a separate server, --mode gguf loads the file in process (basal[gguf], llama-cpp-python): basal-serve --model ./basal-2.0-max-GGUF --mode gguf --gguf basal-2.0-max-GGUF/basal-2.0-max-Q8_0.gguf --port 8000. A plain ollama run gives free text only: the typed answers come from basal-serve in front of it. --http-parallel N (with OLLAMA_NUM_PARALLEL / llama-server -np) serves concurrent clients.

Context. Ollama mode reads at most 4,096 tokens per prompt (basal-serve sends num_ctx 4096, which overrides the Modelfile), and llama.cpp mode reads what llama-server -c allows; longer states are cut in the middle (usage.truncated). For long documents use llama-server -c 32768 or --mode gguf.

Speed: not measured in this release. GGUF is the slow path for batches: it is for desktops and laptops, not servers.

Memory on Apple Silicon. basal-2.0-max-Q8_0.gguf is 12 GB. The max GGUF files were not checked on a Mac: on Apple M4 Pro, 24 GB, Ollama and llama-server ran out of Metal memory while another ~5 GB GPU process was running.

Engine. The same basal engine v2.5.0 also serves the experimental basal-omni-1.0 models (images and audio): README.

LM Studio: basal-2.0-max-Q8_0-decision.gguf and basal-2.0-max-Q5_K_M-decision.gguf are the files above with basal's decision metadata added (weights unchanged). LM Studio 0.4.26 or newer (beta channel at the time of writing; llama.cpp engine 2.65.1+, not the MLX engine) serves them as decision models on /v1/systemone and /v1/decisions: choice, yes/no and score, up to 26 options, temperatures from CALIBRATION.llamacpp.json. LM Studio ends a request after 30 s, so split very long documents.

With basal 2.5.0 you can add the same metadata to a GGUF file yourself (a file of this repository, downloaded, or your own quantisation; the tensors are not touched): basal-export-gguf Remek/basal-2.0-max-GGUF lmstudio/basal-2.0-max-Q8_0.gguf --gguf basal-2.0-max-Q8_0.gguf --systemone.

Calibration

CALIBRATION.json in this folder is the bf16 model's calibration, the default for every mode without a fit of its own. basal-serve --mode ollama and --mode llamacpp load CALIBRATION.<engine>.json for their engine when the folder has it, else CALIBRATION.json. --mode gguf (in process) loads CALIBRATION.json, the bf16 fit; pass --calibration <folder>/CALIBRATION.llamacpp.json to use the llama.cpp fit. A calibration fitted on bf16 readouts does not carry over exactly to a quantised port (an outside test of basal-1.5 on Ollama saw 2.3% real error at the 1% threshold with the bf16 file), so a port-specific fit, where it passed, uses the bf16 model's method on that port's own readouts of the same calibration splits, one option order:

• basal-serve --mode ollama: Ships CALIBRATION.ollama.json, fitted on Ollama (GGUF) readouts of this port (same method as the bf16 fit, one option order); dev v2 ECE 0.0173 with the bf16 fit -> 0.0179 with this fit (in-sample), 0.0187 held-out (leave-one-family-out).
• basal-serve --mode llamacpp: Ships CALIBRATION.llamacpp.json, fitted on llama.cpp (GGUF) readouts of this port (same method as the bf16 fit, one option order); dev v2 ECE 0.0272 with the bf16 fit -> 0.0270 with this fit (in-sample), 0.0285 held-out (leave-one-family-out).

Where a port uses the bf16 calibration, treat its confidence thresholds as approximate and fit your own with basal-calibrate when they matter. CALIBRATION.ollama.json (choice, confidence 0.989) certifies a 1% threshold, the only shipped file that does: with it "max_error": 0.01 on an act or tool question can automate decisions of that type in that mode instead of deferring them all.

Results of the model

See the model card basal-2.0-max for Werdykt, JDI, LEK, PL-Holdout and PL-Real numbers against basal-1.5 and the evaluation notes. This port is not scored separately on those benchmarks.

Limitations

• Validate on your own documents before relying on it: hidden and held-out sets do not cover every domain.
• Polish world knowledge of a small model is limited: put the relevant facts in the state. Legal rules change.
• Exact arithmetic and thresholds are a weak spot of one-pass decisions: compute them in code (or "facts": "auto").
• The calibration is cautious on unseen tasks: of the shipped files only CALIBRATION.ollama.json (choice, confidence 0.989) here certifies a 1% threshold; with the others "max_error": 0.01 on an act or tool question defers every decision; use 0.05, or certify thresholds on your own data with basal-calibrate.
• act and multi over many labels are experimental. Up to 255 options per question (Ollama: 20); the model's context is 32,768 tokens, but Ollama mode reads at most 4,096 tokens per prompt and llama.cpp mode what llama-server -c allows (see Use). The 2.0 models ship no evidence head.
• Decisions with serious consequences for people should be reviewed by a person.

Citation

@software{kinas2026basal20,
  title   = {basal-2.0: typed-decision models and inference engine for Polish and English},
  author  = {Kinas, Remigiusz},
  year    = {2026},
  version = {2.0.0},
  url     = {https://github.com/rkinas/basal}
}

License and attribution

Apache-2.0. Fine-tuned from speakleash/Bielik-PL-11B-v3.0-Instruct (Apache-2.0). Bielik is a project of SpeakLeash and ACK Cyfronet AGH.