nikeshad2017/safecare-ai

SafeCare AI

SafeCare AI is a research prototype for a locally deployable, safety-constrained, knowledge-distilled conversational support model.

Research question: Can a small causal language model retain useful empathetic dialogue behavior from a larger teacher while explicit symbolic safety rules reduce unsafe outputs?

Architecture

Teacher LLM -> teacher responses/top-k logits -> student KD -> symbolic safety -> local CLI/Streamlit

Experimental tracks

A base student; B SFT; C KD; D KD+symbolic safety; E KD+safety+RAG (future).

Measure quality, empathy, safety, escalation correctness, latency, RAM/VRAM and model size. Never invent results; run the evaluation scripts.

Data/licensing

EmpatheticDialogues is optional. Check the current dataset card and institutional rules before redistribution or commercial use. Keep raw data out of git.

Setup

Python 3.10+ recommended.

Windows: python -m venv .venv .venv\Scripts\activate pip install -r requirements.txt

Linux/macOS: python3 -m venv .venv source .venv/bin/activate pip install -r requirements.txt

Install the correct CUDA-enabled PyTorch build from pytorch.org when applicable.

Pipeline

1. python scripts/download_empathetic.py
2. python scripts/prepare_data.py
3. python scripts/generate_teacher_data.py
4. python scripts/train_sft.py
5. python scripts/distill_topk.py
6. python cli.py or streamlit run app.py
7. python tests/test_safety.py
8. python scripts/evaluate.py --model ./models/safecare-student

Safety

This is a research prototype, not a therapist, diagnosis system, medical device, or emergency service. Keyword rules are only a starter defense. A serious deployment needs a validated risk classifier, adversarial testing, multilingual evaluation, privacy controls, human/clinical review, incident response, and formal governance.