The model fine-tunes the full Qwen/Qwen3-Embedding-0.6B encoder with a
1024 -> 256 -> 6 sigmoid head. The exported model.pt is a TorchScript
bundle for DJL integration; text_encoder/ and head.pt are included for
Python inference and reproducibility.
Results
The held-out test split contains 819 records. Mean absolute error is
0.078755 across the six dimensions.
Per-dimension MAE:
Dimension
MAE
valence
0.05160
arousal
0.06291
dominance
0.05799
connectionNeed
0.06470
openness
0.07713
confidence
0.15821
The training corpus was generated and reviewed synthetically. These metrics
are not a substitute for evaluation on consented, representative human data.
The model is not a clinical or mental-health diagnostic tool. Sarcasm,
negation, code-switching, cultural context and ambiguous short messages may
be unreliable. Treat confidence as an uncertainty signal, not a probability
that the prediction is correct.
The raw training corpus is intentionally not included. Provide a local JSONL
corpus with the fields sampleId, text, and the six output names, then run:
python scripts/train-user-affect-model.py --corpus path/to/corpus.jsonl --output model
Training requires CUDA and BF16. CPU fallback is disabled by design.
License and attribution
The repository code and Thymos model artifacts are licensed under
AGPL-3.0-or-later. The model is a fine-tuned derivative of
Qwen3-Embedding-0.6B; upstream copyright, attribution, license notices and
terms continue to apply. See the Qwen model card before redistribution or
commercial use.
