The emotion-recognition backbone reported in the accompanying paper: RoBERTa-large, domain-adaptively pretrained on HIV-contextualized messages, then fine-tuned in two phases (GoEmotions, then GoEmotions merged with the HIV-contextualized data) with focal loss. Seven sigmoid outputs (anger, disgust, embarrassment, fear, grief, nervousness, sadness) — the model has no neutral class and never predicts neutral. In the deployed pipeline the backbone passes its seven probabilities to an LLM (GPT-5-mini, 4-shot), which makes the final single-label decision over eight classes, neutral included. See referenced paper at the end for more details.

Deployed, F1-tuned (best_thr_f1.*) — use these to reproduce the paper: anger 0.45, disgust 0.65, embarrassment 0.85, fear 0.50, grief 0.40, nervousness 0.90, sadness 0.70. A recall-oriented set (β=2, best_thr_recall.*: 0.15/0.40/0.75/0.40/0.20/0.40/0.45) is included for screening-style use; it is not the paper's configuration.

To use or score the backbone without the LLM stage, the paper assigns the highest-probability emotion among those crossing their thresholds, and neutral when none crosses (Table 3.2's backbone row, Appendix E.5). That mapping is an evaluation convention, not a model output — in the deployed system, neutral decisions come from the LLM stage.

DAPT (935 messages, MLM 0.15, seq 256) → phase 1 on GoEmotions (7-label projection) → phase 2 on GoEmotions merged with the HIV data (focal γ=2, label smoothing 0.1, authored-message downweighting). Trained on a single NVIDIA T4 (Google Colab) with gradient accumulation preserving effective batch sizes of 32/32/16; transformers==4.57.1, seed 42.

Not a medical device; research use only. Weakest categories are fear and grief (recall 0.68 in the full pipeline). English only. Evaluation uses authored messages whose intended labels were fixed at authoring time. Single training run; GPU nondeterminism means retraining lands near, not on, these numbers.

To come... <CITATION — paper reference to be added>