File size: 6,188 Bytes
6256b19 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 | """Text2KGBench-style eval of the fine-tuned model on held-out test.jsonl.
Metrics per task family:
IES Turtle : syntactic validity, ontology conformance (our validator), hallucinated-term rate
multi-standard: syntactic validity, relation conformance (predicates within the given ontology)
Usage: python eval_ies.py --model <base> --adapter adapters (omit --adapter for baseline)
Runs under .venv (mlx_lm)."""
import sys, json, argparse, pathlib, re
sys.path.insert(0, str(pathlib.Path(__file__).parent))
from iesval import validate_turtle, structural_conformance, REAL, IES
from rdflib import Graph
from mlx_lm import load, generate
ROOT = pathlib.Path("/Users/fabio/projects/qwen-ies-ft")
def extract_turtle(txt):
m = re.search(r"```(?:turtle|ttl)?\s*(.*?)```", txt, re.S)
if m: txt = m.group(1)
return txt.strip()
def ies_metrics(ttl, user_turn=""):
ok, reason, n, used = validate_turtle(ttl, min_triples=3)
# syntactic validity independent of conformance
syn = True
try: Graph().parse(data=ttl, format="turtle")
except Exception: syn = False
# hallucinated term rate among ies: terms
terms = set(re.findall(r"ies:(\w+)", ttl))
halluc = (len(terms - REAL)/len(terms)) if terms else 1.0
# structural conformance (domain/range with subclass closure)
sc, checked, _ = structural_conformance(ttl) if syn else (0.0, 0, [])
# namespace fidelity: if the prompt demanded a namespace, the output must use it
m = re.search(r"Use <(\S+)> as the namespace", user_turn)
ns_ok = (m.group(1) in ttl) if m else None
return {"syntactic": syn, "conformant": ok, "halluc_rate": halluc,
"struct_conf": sc, "ns_ok": ns_ok, "triples": n}
def multistd_metrics(ttl, allowed_slugs):
syn = True
try: Graph().parse(data=ttl, format="turtle")
except Exception: syn = False
# extract predicates properly via RDF parsing (regex falsely captured subjects)
flat = set()
try:
_g = Graph(); _g.parse(data=ttl, format="turtle")
EXNS = "http://example.org/kg#"
flat = {str(p).split("#")[-1] for _, p, _ in _g if str(p).startswith(EXNS)}
except Exception:
pass
# IES bleed check: an ontology-conditioned multistd answer must not emit ies: terms
ies_bleed = bool(re.search(r"\bies:\w+", ttl))
bad = flat - allowed_slugs
return {"syntactic": syn, "rel_conformant": len(bad)==0, "ies_bleed": ies_bleed,
"off_ontology": sorted(bad)[:5]}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--model", default="mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit")
ap.add_argument("--adapter", default=None)
ap.add_argument("--max", type=int, default=200)
a = ap.parse_args()
kw = {"adapter_path": a.adapter} if a.adapter else {}
model, tok = load(a.model, **kw)
test = [json.loads(l) for l in (ROOT/"data"/"mlx"/"test.jsonl").open()][:a.max]
ood_p = ROOT/"data"/"mlx"/"ood_test.jsonl"
ood_rows = [json.loads(l) for l in ood_p.open()] if ood_p.exists() else []
ies, multi, ood = [], [], []
for r, bucket_ood in [(x, False) for x in test] + [(x, True) for x in ood_rows]:
msgs = r["messages"]
sysp = msgs[0]["content"]
prompt = tok.apply_chat_template(msgs[:-1], add_generation_prompt=True, tokenize=False)
out = generate(model, tok, prompt=prompt, max_tokens=1400, verbose=False)
ttl = extract_turtle(out)
if sysp.startswith("You extract ontology-conformant"):
um = msgs[1]["content"]
rels = re.search(r"Allowed relations: (.+?)\.", um)
allowed = set()
if rels:
for x in rels.group(1).split(","):
allowed.add(re.sub(r"[^0-9A-Za-z]+","_",x.strip()).strip("_"))
multi.append(multistd_metrics(ttl, allowed))
elif "@prefix ies" in msgs[-1]["content"] or "a ies:" in msgs[-1]["content"]:
(ood if bucket_ood else ies).append(ies_metrics(ttl, msgs[1]["content"]))
def pct(rows, key): return 100.0*sum(1 for r in rows if r[key])/max(1,len(rows))
def mean(rows, key): return sum(r[key] for r in rows)/max(1,len(rows))
def block(name, rows):
print(f"=== {name} (n={len(rows)}) ===")
if not rows: return
print(f" syntactic validity : {pct(rows,'syntactic'):.1f}%")
print(f" term conformance : {pct(rows,'conformant'):.1f}%")
print(f" structural conf : {mean(rows,'struct_conf'):.3f}")
print(f" halluc-term rate : {mean(rows,'halluc_rate'):.3f}")
ns = [r for r in rows if r["ns_ok"] is not None]
if ns: print(f" namespace fidelity : {pct(ns,'ns_ok'):.1f}% (n={len(ns)})")
print()
block("IES Turtle (in-distribution)", ies)
block("IES Turtle (OUT-OF-DISTRIBUTION, gold dstl)", ood)
print(f"=== Multi-standard (n={len(multi)}) ===")
if multi:
print(f" syntactic validity : {pct(multi,'syntactic'):.1f}%")
print(f" relation conformance: {pct(multi,'rel_conformant'):.1f}%")
print(f" IES-bleed rate : {pct(multi,'ies_bleed'):.1f}% (should be 0)")
out = {
"ies_n": len(ies), "ood_n": len(ood), "multi_n": len(multi),
"ies_syn": f"{pct(ies,'syntactic'):.1f}%" if ies else "-",
"ies_conf": f"{pct(ies,'conformant'):.1f}%" if ies else "-",
"ies_struct": f"{mean(ies,'struct_conf'):.3f}" if ies else "-",
"ies_hall": f"{mean(ies,'halluc_rate'):.3f}" if ies else "-",
"ood_syn": f"{pct(ood,'syntactic'):.1f}%" if ood else "-",
"ood_conf": f"{pct(ood,'conformant'):.1f}%" if ood else "-",
"ood_struct": f"{mean(ood,'struct_conf'):.3f}" if ood else "-",
"multi_syn": f"{pct(multi,'syntactic'):.1f}%" if multi else "-",
"multi_rel": f"{pct(multi,'rel_conformant'):.1f}%" if multi else "-",
"multi_bleed": f"{pct(multi,'ies_bleed'):.1f}%" if multi else "-",
"adapter": a.adapter or "baseline",
}
(ROOT/"data"/("eval.json" if a.adapter else "eval_baseline.json")).write_text(json.dumps(out, indent=2))
print("wrote", "eval.json" if a.adapter else "eval_baseline.json")
if __name__=="__main__":
main()
|