Jevify — Mintzs

Partly. The 250 single-decision cases ran on one RTX 3050 laptop GPU with the same checkpoint, and some cases were used during prompt tuning. Single-decision requests also do not show multi-question caching gains.

No. It wraps an existing checkpoint with prompting, scoring, and execution changes, and leaves weights unchanged. It explicitly does not reproduce the Jev architecture or RLCD training.

Only specific dense Qwen2, Qwen3, and Llama-style checkpoints with compatible architecture and tokenizer. Mixture-of-experts, GGUF/Ollama endpoints, quantized wrappers, and multi-GPU setups are unsupported.

Roughly 3 GB of weights and tokenizer from Hugging Face on first use, cached locally and reused afterwards. Set HF_HOME to choose the folder, or pass --local-files-only to require cached files instead of downloading again. Before that you need Python 3.11+, a CUDA-enabled PyTorch build for NVIDIA GPUs, then pip install -e ".[inference]". The default is Qwen/Qwen2.5-1.5B-Instruct and CPU execution is supported.

Correct plus schema-valid needs both the right choice and valid, complete JSON including probabilities for every option, so formatting failures count as misses. That is why standard Qwen scores 9.2% while Jevify reaches 90.0%; the plain model's 9.2% is not its standalone classification accuracy.

Because they were slower here. The 167 ms configuration used optional RMSNorm, SwiGLU, and RoPE kernels with CUDA graphs disabled, while the graphs-enabled row took 951 ms on the same 250 cases, since capture overhead outweighed the gain for these shapes. CUDA graphs remain an option for repeated identical input shapes.

Edit workflow.json, which holds three question types: choice for named options, noul for a yes/no probability, and score for an ordered rubric. The included workflow demonstrates refund triage; pass --workflow path/to/workflow.json to switch, and supply fresh context or conversation history with each request.