Vibe Coding DiscoverSemIf
View on GitHubSemantic ifs from open models, on a 3090 at home. Independent; not affiliated with Jev or TypeSafe.
SemIf reads typed option probabilities directly from open LLM logits to answer runtime-defined decisions (routing, retries, evidence checks) without generating or parsing text. Includes Python CLI, frozen benchmarks, and a WebGPU browser demo on 4B-class models.
Use Cases
Score runtime-defined yes/no or multi-option decisions from model logitsRoute support tickets into queuesDecide whether evidence supports a claim (NLI)Let agents make small routing/retry decisions without generating textRerank retrieval candidates for RAGBenchmark decision accuracy vs autoregressive JSON baselinesRun quantized models in-browser for decision demosReuse a shared prompt state across many criteria for faster scoring
Built With
- Language
- Python
- Frameworks
- PyTorch · Hugging Face Transformers · Accelerate · safetensors · WebGPU · llama.cpp/GGUF · NumPy
Tags
semantic-decisions · logit-readout · llm-inference · classification · benchmarking · agents · webgpu · gguf · quantization · evaluation · probability-calibration · reranking · pytorch · reproducibility · local-gpu · runtime-criteria