raghunath1/Aptivra-Base-110M

Card: v91soup40 backend-parity table

Add v91soup40 OpenVINO fp32 (0.99985)

Add v91soup40 ONNX (fp32 1.00000, O3 0.99999, int8 0.99422)

Coherence fix: sync repo to v91soup40 release exactly (add tokenizer/2_Normalize; remove stale v5 onnx/openvino binaries — regeneration pending).

Release v91soup40: boundary-MNRL soup. Prod human R@1 0.99, adversarial 0.981, near-twin boundary R@1 0.42->0.70. safetensors only (ONNX/OpenVINO/GGUF regen pending).

docs(card): OpenVINO int8 (weight-only) validated + added

add OpenVINO int8 (weight-only, fidelity 0.99938)

add OpenVINO int8 (weight-only, fidelity 0.99938)

docs(card): final backend parity table (ONNX/OpenVINO/GGUF validated; OV int8 withheld)

add OpenVINO fp backend — parity-validated (int8 withheld pending recalibration)

add ONNX backends (fp, O3, int8 x4) — parity-validated

docs(card): best-in-class multi-backend rewrite (ONNX/OpenVINO/GGUF), per-OS usage, bare-text input fix, architecture diagram; interim parity table

add assets/aptivra-architecture.svg

add assets/aptivra-architecture.png

docs(card): neutralize human-in-loop wording; tighten R@1 to exact 0.958; add rerank-ON fixture-coupling qualifier (card-only, weights untouched)

Model card: neutral experimental scope wording

Honest model card: experimental scope, base-lift vs e5 (CI), reranker not in artifact, expert-review required

Set MIT license

Upload Aptivra-Base-110M skill embedding model

initial commit