zer-judge 1.0.0

ONNX-based neural judge for zer, runs DeBERTa/MiniLM NLI models via ORT to adjudicate borderline record pairs
Documentation

zer-judge

ONNX-based neural judge for the zer entity-resolution library.

Loads a DeBERTa-v3 or MiniLM NLI cross-encoder model via ONNX Runtime and uses it to adjudicate borderline record pairs that the Fellegi-Sunter scorer could not classify with high confidence. Models are hosted on HuggingFace at arsalan-anwari/zjudge.

Getting the models

# Download all model variants to ~/.cache/zer/models/
bash scripts/download_models.sh

# Or point at a custom location:
export ZER_MODEL_DIR=/path/to/your/models

The ZER_MODEL_DIR environment variable controls where zer looks for models. If unset, zer checks ~/.cache/zer/models and falls back to ./models.

Feature flags

Flag Description
judge_cpu CPU execution provider for ORT
judge_cuda NVIDIA CUDA execution provider (requires CUDA toolkit + cuDNN)
judge_tensorrt NVIDIA TensorRT EP, FP16 + engine caching (requires TensorRT 8+)
judge_rocm AMD ROCm execution provider
judge_directml Windows DirectML execution provider
judge_openvino Intel OpenVINO execution provider

These are independent from zer-compute's cuda/avx2/vulkan flags.

Usage

[dependencies]
zer = { version = "1.0", features = ["judge_cpu"] }
# or directly:
zer-judge = { version = "1.0", features = ["judge_cpu"] }

use zer_judge::{spec_from_env, ModelPrecision, backend::JudgeBackend};

// Reads ZER_MODEL_DIR → ~/.cache/zer/models → ./models
let backend = JudgeBackend::auto_detect();
let spec    = spec_from_env(ModelPrecision::Fp16Fused, backend.available_vram_bytes());

// Or pick a specific model explicitly:
use zer_judge::spec::DebertaBaseSpec;
let spec = DebertaBaseSpec::from_env(ModelPrecision::Base);

Part of the zer ecosystem

zer · GitHub