Checking for a trained model…
🤗 EmpaChat
GPT-2 XL with two LoRA adapters: one reads the emotion in a message, the other writes a reply the way a supportive listener would. Every number and reply on this page is read live from the training run's output on the Hugging Face Hub.
Measured results
From metrics.json, written by the training run itself.
—
Emotion classifier · test accuracy
—
Classifier · macro-F1 (3-class)
—
Reply generator · validation perplexity
—
LoRA · trainable parameters
Full classification report
How it works
Two fine-tuned copies of GPT-2 XL (1.5B parameters), each trained with 4-bit QLoRA.
Your message
Normalised the same way as the training data.
Emotion classifier
Trained on GoEmotions Reddit comments → positive · negative · neutral.
Reply generator
Trained on real listener replies from EmpatheticDialogues, conditioned on that emotion.
Replies from the trained model
Generated by the reply generator at the end of training (samples.json), unedited.
Waiting for the first training run to publish its samples.