SurvAIv Ettin 32M — v3 listwise, epoch 1

SurvAIv's fine-tuned Ettin reranker. This is a complete modular Sentence Transformers CrossEncoder checkpoint, not a LoRA adapter. Keep every numbered module directory: the base transformer weights alone omit the trained scoring head. Tokenizer and model configuration are included at the repository root.

onnx/reranker-int8.onnx is the INT8 deployment export used in the local SurvAIv RAG. It uses the accompanying tokenizer and a maximum input sequence length of 512. Scores express ranking relevance, not calibrated probabilities of answer correctness or FAQ applicability.

Training: LambdaLoss@10 on 2,098 query groups and 23,657 graded passages. The validation set has 200 groups, of which 142 are evaluable. Training data is hosted separately in the private SurvAIv reranker training dataset. See TRAINING_MODEL_CARD.md for the original generated card, architecture and historical training metrics. Those metrics are not the canonical or Independent-200 RAG benchmark results.

The original card records base revision b33e5ceb5110773ea9cf5e00c9bedc83a8c2afdd. The upstream model is Apache-2.0 licensed. See LICENSE and NOTICE.md. This public release retains the original epoch-1 model weights, not a model retrained on the later -fix dataset. The training dataset remains private.

The Apache-2.0 model licence does not relicense third-party document excerpts in the archived TRAINING_MODEL_CARD.md. Those historical examples have separate source-text conditions, including non-commercial restrictions; see SOURCE_TEXT_NOTICES.md. They are not an unrestricted training-data release. The underlying Ettin encoder's MIT notice is retained in ETTIN_LICENSE.

This is a retrieval component, not a clinically validated first-aid system.