title: SAMS Qwen v7 Approved-Wording Evaluator
emoji: 🏔️
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
suggested_hardware: l40sx1
startup_duration_timeout: 1h
fullWidth: true
SAMS Qwen3-1.7B approved-wording selector v7
Docker Space for an NF4 QLoRA fine-tune of the official Qwen/Qwen3-1.7B base on one NVIDIA L40S. Training uses the official safetensors checkpoint, merges the adapter into the official base, and only then exports GGUF Q4_K_M and Q5_K_M.
V7 packages the trained bounded wording selector with a mandatory deterministic candidate guard; it is not a medical or robot safety controller. Each request supplies trusted facts and four already-approved, grounded natural paraphrases. Qwen selects presentation style only; unsafe or factually conflicting wording is never offered as an option. Deterministic code supplies the question, workflow action, routing flags, and cited facts, then renders the final response. Low-confidence speech, fixed emergency warnings, medical triage, emergency/descent decisions, and robot motion, navigation, motor, joint, power, or shutdown control bypass Qwen.
The bundled dataset has 12,400 source-traceable synthetic conversations: 10,080 train, 1,148 validation, and 1,172 development-test. Raw PPG, ECG, and IMU arrays are not serialized into prompts. HealthBench and MedSafetyBench are excluded from training. Generic robot examples are explicitly not real Unitree G1 expedition logs.
The fresh v7 release gate is frozen before evaluation and excluded from development evaluation. It contains exactly 40 cases. Both Q4_K_M and Q5_K_M must pass all 40; 39/40 is rejected. JSON-schema decoding constrains valid supplied IDs and shape but does not force the expected choice.
The targets remain source-grounded synthetic drafts requiring domain review. Passing this software-contract gate is not medical-device, mountaineering safety, clinical, or robot-safety validation.