Most chatbots reach for solutions. You say something difficult and they suggest meditation, therapy, five actionable steps. Emotwen does the opposite — it listens, reflects, and validates. It does not give advice.
Emotwen is Qwen 3.5 (0.8B parameters) fine-tuned into an empathetic journal companion. The name is Qwen backwards — the model’s job is to mirror, not push forward. At 0.8B, it runs on modest hardware, which matters for a companion that should stay personal.
The “no advice” constraint is not a prompt instruction. It is baked into the training data. An advice-detection filter drops examples like “you should try meditation” and keeps “that sounds really hard.” The model learned restraint because it never saw advising rewarded.
Training is two-stage SFT: first conversational tone from empathetic dialogues, then journal-domain specialization at a lower learning rate. If evaluation shows verbosity creeping in, a GRPO reinforcement step tightens responses to 2-5 sentences. The full pipeline runs in Colab notebooks with Weights & Biases tracking.
Emotwen is an experiment in subtraction — training a model to do less, and seeing whether restraint makes it more useful.
Source: afonsomota/emotwen-3.5-finetune