aigent-stage2-burnout-multilabel-v1

Re:Mind Stage 2 Burnout Signals v1

This is a locally trained, six-label Korean text classification artifact for Re:Mind Stage 2. It is an informational pattern signal model and is not a medical diagnostic model.

Labels

The model uses independent sigmoid outputs for:

exhaustion, overload, helplessness, low_efficacy, anxiety, and irritability.

The probabilities do not sum to one. Use thresholds.json and the per-label calibration status before turning a probability into a signal.

Training and evaluation

• Loss: masked BCE with logits
• Human training rows: 179 group-safe rows
• Negative-only partial weak rows: 4,524
• Independent validation rows: 100
• Weak-label sample weight: 0.10
• Max sequence length: 256
• Model status: shadow_only

The independent validation rows were excluded from training. This artifact is not approved for diagnosis, treatment, automated risk scoring, or eligibility decisions.

Current calibration gate

Only anxiety currently passes the configured precision gate. The remaining labels are marked blocked in thresholds.json and must remain unvalidated in product output.

Files

• model/: Hugging Face model files
• tokenizer/: tokenizer files
• thresholds.json: per-label threshold and calibration status
• label_mapping.json: exact label order and IDs
• run_config.json: training provenance and contract metadata

The serving application must load this artifact locally with local_files_only=True, and must keep the response fields informational_only=true and risk_score_eligible=false.