Safin-1.5

Download all model files:

pip install -U huggingface_hub
hf download AI45Research/Safin-1.5 --local-dir ./Safin-1.5

For environment setup, see Safin-1.

PKG=/path/to/MARCH-Safin-1 MODEL=/path/to/model \
CUDA_VISIBLE_DEVICES=0 bash infer.sh

# In another terminal:
python3 request.py --prompt "Hello" --output response.json

Inference uses repetition_penalty=1.05.

Formal reasoning harness

The optional Safin reasoning harness provides a manual use_harness switch, independent thinking, machine-verified formal reasoning, ProverQA proof preservation and independent rescue, and explicit model fallback.

Install the tool from this repository (Python 3.11+):

pip install -e 'harness[server]'
safin-harness-install-vampire

With the model API already running at http://127.0.0.1:28080/v1, start:

safin-harness-serve --upstream http://127.0.0.1:28080/v1 \
  --model Safin-1.5 --port 28081 --output-root ./harness_outputs

Send requests to the harness endpoint at port 28081. Ordinary chat defaults to use_harness=false. To enable the ProverQA portfolio, set use_harness=true, harness_profile="proverqa", and thinking=true or false. The examples include complete request bodies.

For FOLIO/PrOntoQA bridge models, installation of the optional bridge extra and safin-harness-prepare-models are described in the tool README. The tool is published under Apache-2.0; see its license, third-party notices, and validation scope. This publication adds the tool without claiming new Safin-1.5 evaluation results.