Train domain: hipaa — structured opus-authored PRTs (format isolated).
Cross-domain holdout: evaluate on the OTHER domain's held-out test set.
Prompt format: no-policy zero-shot chain-of-thought; the model reasons step-by-step and ends with Final Verdict: COMPLIANT or Final Verdict: NONCOMPLIANT.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("josephimperial/qwen2.5-7b-prt-hipaa-structured-full-policy")
model = AutoModelForCausalLM.from_pretrained("josephimperial/qwen2.5-7b-prt-hipaa-structured-full-policy", torch_dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": PROMPT}]  # see prepare_data.build_prompt