| """Multi-standard generalization slice from Text2KGBench (Wikidata-TekGen). |
| Builds ontology-CONDITIONED pairs: the prompt supplies the target ontology's relations, |
| the target Turtle uses ONLY those relations (conformance, parallel to the IES validator). |
| Teaches the general 'conform to the ontology you are given' skill. No LLM; rdflib-checked.""" |
| import json, pathlib, re, random, glob |
| from rdflib import Graph |
| random.seed(7) |
| ROOT = pathlib.Path("/Users/fabio/projects/qwen-ies-ft") |
| BASE = ROOT/"Text2KGBench"/"data"/"wikidata_tekgen" |
| PER_DOMAIN = 45 |
| SYS = ("You extract ontology-conformant RDF/Turtle knowledge graphs from text. " |
| "You use only the relations of the target ontology you are given, and you never " |
| "invent relations outside it.") |
|
|
| def slug(s): |
| s = re.sub(r"[^0-9A-Za-z]+", "_", s.strip()).strip("_") |
| return s or "x" |
| DATEISH = re.compile(r"\d{1,2}\s+\w+\s+\d{4}|\d{4}(-\d\d-\d\d)?$|^\d[\d.,]*$") |
|
|
| def to_turtle(triples, rel_slugs): |
| by_sub = {} |
| used_ok = True |
| for t in triples: |
| r = slug(t["rel"]) |
| if r not in rel_slugs: |
| continue |
| by_sub.setdefault(slug(t["sub"]), []).append((r, t["obj"])) |
| if not by_sub: return None |
| lines = ["@prefix ex: <http://example.org/kg#> .", ""] |
| for sub, pos in by_sub.items(): |
| parts = [] |
| for r, obj in pos: |
| if DATEISH.match(obj.strip()): |
| parts.append(f'ex:{r} "{obj.strip()}"') |
| else: |
| parts.append(f"ex:{r} ex:{slug(obj)}") |
| lines.append(f"ex:{sub} " + " ;\n ".join(parts) + " .") |
| return "\n".join(lines) |
|
|
| def main(): |
| onts = {} |
| for f in glob.glob(str(BASE/"ontologies"/"*.json")): |
| d = json.load(open(f)) |
| onts[d["id"]] = d |
| pairs = [] |
| for gt in sorted(glob.glob(str(BASE/"ground_truth"/"*.jsonl"))): |
| oid = "ont_" + re.search(r"ont_(\d+_\w+?)_", pathlib.Path(gt).name).group(1) |
| |
| ont = next((o for k,o in onts.items() if k==oid or k.replace("ont_","")==oid.replace("ont_","")), None) |
| if not ont: continue |
| rels = [r["label"] for r in ont["relations"]] |
| rel_slugs = {slug(r) for r in rels} |
| concepts = [c["label"] for c in ont["concepts"]][:20] |
| rows = [json.loads(l) for l in open(gt)] |
| random.shuffle(rows) |
| kept = 0 |
| for row in rows: |
| if kept >= PER_DOMAIN: break |
| ttl = to_turtle(row.get("triples",[]), rel_slugs) |
| if not ttl: continue |
| try: Graph().parse(data=ttl, format="turtle") |
| except Exception: continue |
| user = (f"Target ontology '{ont['title']}'.\n" |
| f"Allowed relations: {', '.join(rels)}.\n" |
| f"Example concepts: {', '.join(concepts)}.\n\n" |
| f"Extract a knowledge graph from this sentence as RDF/Turtle, using ONLY the " |
| f"allowed relations and the ex: <http://example.org/kg#> namespace. Output only Turtle.\n\n" |
| f"Sentence: {row['sent']}") |
| pairs.append({"messages":[{"role":"system","content":SYS}, |
| {"role":"user","content":user}, |
| {"role":"assistant","content":ttl}]}) |
| kept += 1 |
| out = ROOT/"data"/"pairs_multistd.jsonl" |
| with out.open("w") as f: |
| for p in pairs: f.write(json.dumps(p)+"\n") |
| print(f"multi-standard pairs: {len(pairs)} across {len(onts)} ontologies -> {out}") |
|
|
| if __name__=="__main__": |
| main() |
|
|
