| """ub_d3_common.py (UBD3, 16 Oct 2026): one rendering of a typed decision (Decision Index wire format: state + {type, instructions, |
| criteria}) into the UBENC row contract, shared by the d3 data prep (training rows) and the Decision Index engine (inference), so |
| train and test see byte-identical inputs. |
| |
| state -> one part, role "user": the string as is, or a JSON object as compact JSON (the kit's base.text rendering); |
| instructions -> the question text (string as is, object as compact JSON); |
| choice -> options {key: description or ""} in criteria order (the encoder reads "key: description", or "key" alone); |
| noul -> a 2-option choice {"no": "", "yes": ""}; the answer is p(yes). |
| The question is marked nodrop: training never drops options (d3 trains on the full option set, only re-permuted), and has no [ABS].""" |
| import json |
|
|
| NOUL_OPTIONS = {"no": "", "yes": ""} |
|
|
|
|
| def text(x): |
| return x if isinstance(x, str) else json.dumps(x, ensure_ascii=False, separators=(",", ":")) |
|
|
|
|
| def options_of(q): |
| if q["type"] == "noul": |
| return dict(NOUL_OPTIONS) |
| return {str(k): ("" if v is None else text(v)) for k, v in q["criteria"].items()} |
|
|
|
|
| def to_row(rid, state, q, gold=None, soft=None, weight=1.0): |
| """state: str | object; q: {type, instructions, criteria?}. gold: option name (noul: "yes"/"no"); soft: {option name: prob}.""" |
| opts = options_of(q) |
| qq = dict(id="d3:" + q["type"], type="choice", over="user", text=text(q.get("instructions", "")), question_paraphrases=[], |
| options=opts, abstain=False, nodrop=True, gold=gold if gold is not None else next(iter(opts))) |
| if soft is not None: |
| qq["soft"] = soft |
| if weight != 1.0: |
| qq["weight"] = weight |
| return dict(id=rid, parts=[dict(role="user", text=text(state if state is not None else ""))], questions=[qq]) |
|
|
