obx0x3/empathy-dementia

Empathy Dementia

A fine‑tuned multilingual T5‑Base model for empathetic responses designed for individuals with dementia.

Training Corpus

• Data splits:
  • Training: 120 examples
  • Validation: 30 examples
  • Test: 30 examples
• Languages: English ("en") and French ("fr")
• Data structure: Each sample includes:

  {
    "input": "Where are my glasses?",
    "target": "Let’s look together. Maybe near your chair.",
    "lang": "en"
  }
  

Training Results (Cross-Entropy Loss)

Epoch	Training Loss	Validation Loss
1	0.191100	0.101121
2	0.119100	0.061524
3	0.103100	0.042741
4	0.079900	0.038426

Usage:

from transformers import T5Tokenizer, T5ForConditionalGeneration

model = T5ForConditionalGeneration.from_pretrained("obx0x3/empathy-dementia")
tokenizer = T5Tokenizer.from_pretrained("obx0x3/empathy-dementia")

def empathetic_response(prompt, lang="en"):
    prefix = "emotion: " if lang == "en" else "émotion: "
    input_ids = tokenizer(prefix + prompt, return_tensors="pt").input_ids
    output_ids = model.generate(input_ids, max_length=50)
    return tokenizer.decode(output_ids[0], skip_special_tokens=True)

Example

print(empathetic_response("I feel so lonely.", lang="en"))
print(empathetic_response("Je me sens seul.", lang="fr"))

Intended Use & Limitations

MSc Dissertion Proof of concept and work for primary use case: Aid communication with individuals experiencing memory loss, confusion, or distress

NOT for clinical decisions — designed as a supportive tool