Bielik-One-4B-r1-Adapter

vLLM package of the Qwen3.5 4B decision model: LoRA adapter + separate decision heads, loaded by the heads plugin on the unchanged base Qwen/Qwen3.5-4B. The same model as a single merged checkpoint: Bielik-One-4B-r1.

Small add-on to an existing deployment. If you already serve Qwen/Qwen3.5-4B with vLLM, this package adds only about 0.7 GB of GPU memory: the LoRA adapter and the decision heads. The base weights are shared, so one engine keeps answering ordinary chat with the unchanged base model and serves typed decisions next to it. (A merged 4B model would need its own full copy of about 9 GB.)

Start

pip install "bielik-one[vllm] @ git+https://github.com/speakleash/bielik-one"
pip install --no-deps "systemone-vllm-heads @ git+https://github.com/speakleash/bielik-one#subdirectory=vllm-heads-plugin"
hf download SPKL-Internal/Bielik-One-4B-r1-Adapter --local-dir model
bielik-one backend model    # first terminal; downloads the pinned base if needed
bielik-one frontend model   # second terminal

Typed decisions: http://127.0.0.1:8012/v1/systemone, model decision. Original base chat: http://127.0.0.1:8012/v1/chat/completions, model qwen-base; the base weights stay unchanged and are shared with the adapter on one GPU. Do not use decision for free-text generation. Default profile: vLLM 0.32.0, one fixed adapter, bfloat16 base weights (the precision of the reported results and of the calibration), FP32 readouts, prefix caching, context 32768, 32 sequences, 4096 batched tokens; --fp8 loads FP8 base weights to save GPU memory, changes some decisions near the decision boundary and is not calibrated. GPU memory in bfloat16: about 12 GB for this 4B model (weights plus the configured KV cache); --fp8 roughly halves the weight memory.

Request example

One state, all four question types:

curl -s localhost:8012/v1/systemone -H 'Content-Type: application/json' -d '{
  "model": "decision",
  "state": "Invoice 12/2026 for EUR 4,300 has not been paid. The payment deadline passed yesterday. The customer has paid late twice this year.",
  "questions": {
    "status": {
      "type": "choice",
      "instructions": "What is the status of the invoice?",
      "criteria": {
        "paid": "Paid",
        "overdue": "Unpaid and past the deadline",
        "not_due": "Unpaid, deadline not reached"
      }
    },
    "late": {
      "type": "noul",
      "instructions": "Has the payment deadline passed?",
      "criteria": {
        "true": "Yes",
        "false": "No"
      }
    },
    "urgency": {
      "type": "score",
      "instructions": "How urgent is collection?",
      "criteria": [
        "none",
        "low",
        "medium",
        "high",
        "critical"
      ]
    },
    "actions": {
      "type": "labels",
      "instructions": "Which actions are appropriate now?",
      "criteria": {
        "reminder": "Send a payment reminder",
        "interest": "Charge late-payment interest",
        "cancel": "Cancel the invoice"
      }
    }
  }
}'

Answers, keyed by question name: choice gives the chosen key, probabilities per option and confidence; noul gives the probability of yes; score gives the expected level (score, 0-based over the listed levels), probabilities per level, the legend and confidence; labels gives an independent probability per option and the selected keys (probability >= 0.5). Interactive schema: /docs.

Calibration

calibration.json holds one temperature per question type, fitted on the validation split only: choice T=1.625, binary T=1.625, score T=2.878, labels T=0.917. Categorical: p = softmax(log p_raw / T); labels: p = sigmoid(logit(p_raw) / T) per option. Calibration does not change argmax or 0.5-threshold decisions. bielik-one applies it by default for the standard request (one option order); --no-calibration or "temperature": 1 returns raw probabilities. It was fitted on bfloat16 outputs and is not fitted for --fp8 serving.

Data Protection and Copyright Requests

For requests concerning personal data or copyrighted content, contact biuro@speakleash.org.pl. Please include the model name and details needed to identify the material concerned.

Citation

If you use this model, please cite this model card and acknowledge the underlying base model:

@misc{wrobel2026BielikOne4Br1Adapter,
  author  = {Wróbel, Krzysztof and Ociepa, Krzysztof and Flis, Łukasz and Gwoździej, Adrian and Jezierski, Karol and Ciuciura, Igor and {SpeakLeash Team} and {Cyfronet Team}},
  title   = {Bielik-One-4B-r1-Adapter},
  year    = {2026},
  version = {r1},
  url     = {https://huggingface.co/SPKL-Internal/Bielik-One-4B-r1-Adapter},
  note    = {Bielik-One decision model: model card and weights}
}

Credits

• Krzysztof Wróbel^SpeakLeash - decision-model design, training, evaluation and inference implementation
• Krzysztof Ociepa^SpeakLeash - consultations
• Qwen Team - development of the underlying Qwen3.5-4B base model.

We thank the entire SpeakLeash team for their commitment to the Bielik project and the collaboration between the open-science SpeakLeash project and the HPC center ACK Cyfronet AGH. We gratefully acknowledge the contributions of: Sebastian Kondracki, Marek Magryś, Sonia Staniek, Waldemar Boszko, Maria Czarkowska, Jacek Chwiła, Kuba Sołtys, Dominika Basaj, Remigiusz Kinas, Marcin Dąbrowski, Anna Przybył, Agnieszka Ratajska, Daria Bartkowiak, Witold Wydmański, Jan Sowa, Grzegorz Anioł, Bartosz Bartnicki, and many other wonderful researchers and enthusiasts of the AI world.

Contact Us

If you have any questions or suggestions, please use the discussion tab. If you want to contact us directly, join our Discord SpeakLeash.

Acknowledgements

We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Center: ACC Cyfronet AGH) for providing computer facilities and support within computational grant no. [PLG/2026/019559].

This work was supported by the Gaia AI Factory project, funded by the European Union under Grant Agreement No. 101314359 through the EuroHPC Joint Undertaking (EuroHPC JU).