scripts/build_refusals.py

"""Refusal/boundary pairs: prompts asking for things IES4 cannot express (subjective
qualities, speculation, opinions, non-RDF outputs). Teaches the model to explain the
limit and offer the nearest representable alternative instead of hallucinating a graph."""
import json, pathlib, random
random.seed(31)
ROOT = pathlib.Path("/Users/fabio/projects/qwen-ies-ft")
SYS = ("You are an expert in the UK Government Information Exchange Standard (IES4), "
       "a 4D RDF ontology. You explain IES concepts precisely and write valid IES4 "
       "RDF/Turtle using only real IES4 terms.")

CASES = [
 ("Record that Fred Smith is a kind and trustworthy person.",
  "IES4 models objective, exchangeable facts (entities, states, events, periods, identifiers), "
  "not subjective character judgements like kindness or trustworthiness. I can encode Fred Smith "
  "as an ies:Person with names, identifiers, states and event participations. If an authority "
  "issued a formal assessment, that assessment event itself could be modelled (who assessed, "
  "when), but the subjective quality cannot be asserted as a fact. Shall I encode the person "
  "record or an assessment event instead?"),
 ("Encode that Alice will probably move to Manchester next year.",
  "IES4 asserts facts about states of the world, not probabilities or speculation about the "
  "future. 'Probably moves next year' is a prediction, not an exchangeable fact. What IES4 can "
  "represent: a possible-world construct for hypotheticals (ies:PossibleWorld) used in assessment "
  "contexts, or a recorded statement/assessment event in which an analyst asserted this "
  "prediction on a given date. Tell me which and I will encode it."),
 ("Write the IES4 Turtle showing that the merger caused the redundancies.",
  "IES4 has no causal relation between events; it records what happened, who participated, when "
  "and where, but not that one event caused another. I can encode both events with their periods "
  "and participants, and, if an analyst made a causal assessment, the assessment itself as an "
  "event. Should I do that?"),
 ("Give me the IES4 representation of Fred's opinion that the policy is bad.",
  "Opinions are not exchangeable facts in IES4. What can be modelled is the communication act: "
  "Fred issuing a statement at a time, via a channel, with the statement content held as a "
  "representation. The truth or sentiment of the opinion is out of scope. Want the communication "
  "encoding?"),
 ("Encode Fred Smith's IQ of 140 and his star sign in IES4.",
  "A measured value like an IQ score can be attached via the characteristics/measure pattern if "
  "you need it, but a star sign is an astrological label, not a fact IES4 exchanges. I can encode "
  "the person, identifiers, and any objectively measured characteristics with their measurement "
  "provenance. Confirm which measured characteristics you want included."),
 ("Output this scenario as SQL INSERT statements using IES table names.",
  "IES4 is an RDF Schema; its conformant serialisations are RDF formats (Turtle, JSON-LD, "
  "N-Triples), not SQL tables. I can produce the scenario as IES4 Turtle, which you can load "
  "into a triple store. Shall I proceed with Turtle?"),
 ("Model that the organisation is evil in IES4.",
  "Moral judgements are outside IES4's scope. I can model the organisation, its posts, assets, "
  "events it participated in, and formal assessments made about it by named assessors at given "
  "times. Which facts should I encode?"),
 ("Encode that Fred loves Alice in IES4 Turtle.",
  "Emotional states are not part of IES4's exchangeable model. What IES4 can express: a "
  "relationship state between two people over a period (for formalised relationships such as "
  "marriage or partnership via states of the persons), or communication events between them. "
  "Tell me which representable relationship you mean and I will encode it."),
]

VARIANTS = [
 "As IES4 Turtle: {q}", "{q} Output only Turtle.", "Using the IES ontology, {q}",
 "{q}", "For our knowledge graph, {q}",
]

pairs = []
for q, a in CASES:
    for v in random.sample(VARIANTS, 5):
        pairs.append({"messages":[
            {"role":"system","content":SYS},
            {"role":"user","content":v.format(q=q[0].lower()+q[1:]) if v!="{q}" else q},
            {"role":"assistant","content":a}]})

out = ROOT/"data"/"pairs_refusal.jsonl"
with out.open("w") as f:
    for p in pairs: f.write(json.dumps(p)+"\n")
print(f"refusal pairs: {len(pairs)} -> {out}")