sam-qwen3-1.7b-sams-v7

SAMS Qwen3-1.7B approved-wording selector v7

This release was trained from the official Qwen/Qwen3-1.7B safetensors checkpoint using 4-bit NF4 QLoRA with BF16 compute. The adapter was merged into the official base before GGUF Q4_K_M and Q5_K_M quantization.

Strict frozen release gate: Q4_K_M 40/40; Q5_K_M 40/40. Every supplied wording is prevalidated for grounding and safety before inference; the model chooses style among four approved paraphrases. The JSON schema allows all supplied IDs and does not contain a hidden expected answer. A score of 39/40 is rejected.

The model selects one supplied approved natural-language response. The deterministic controller supplies the question, workflow action, routing flags, and cited facts. Low-confidence speech, fixed warnings, medical triage, emergency/descent decisions, and robot motion, navigation, motor, joint, and power control bypass this model.

Important: the targets are source-grounded synthetic drafts requiring domain review. The 40-case result is a bounded software-contract regression result, not clinical, medical-device, mountaineering-safety, or robot-safety validation. Generic robot fields are not real Unitree G1 Himalayan logs.