File size: 4,965 Bytes
68f8a08 | 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 | #!/usr/bin/env python3
"""
Validator and metrics for space-ontology model output.
Metrics, chosen to be checkable rather than impressive:
parse_rate fraction of Turtle outputs rdflib can parse
term_conformance fraction of outputs where EVERY ssao: term exists in SSAO
hallucinated_rate invented ssao: terms per output (the headline number)
namespace_fidelity fraction declaring the correct SSAO namespace IRI
class_accuracy fraction whose primary type matches the derived gold class
regime_accuracy fraction whose orbit class matches the derived gold regime
refusal_precision on questions the catalogue cannot answer, did the model refuse
"""
import json
import pathlib
import re
from rdflib import Graph
ROOT = pathlib.Path("/Users/fabio/projects/qwen-space-ft")
SSAO = "https://purl.org/space-ontology/"
VOCAB = json.loads((ROOT / "data" / "vocab.json").read_text())
REAL = set(VOCAB["all_terms"])
REFUSAL_MARKERS = ("cannot determine", "cannot be established", "unknown",
"does not record", "would be fabrication", "no orbit class applies",
"cannot discriminate", "not determined", "insufficient")
def strip_fence(t: str) -> str:
t = re.sub(r"^```[a-z]*\n?", "", t.strip(), flags=re.M)
return re.sub(r"```$", "", t, flags=re.M).strip()
def ssao_terms(text: str):
return set(re.findall(r"ssao:([A-Za-z0-9_]+)", text))
def score_turtle(output: str, gold: str):
out = strip_fence(output)
res = {"parses": False, "terms_ok": False, "hallucinated": [], "ns_ok": False,
"class_match": None, "regime_match": None}
res["ns_ok"] = SSAO in out
used = ssao_terms(out)
bad = sorted(used - REAL)
res["hallucinated"] = bad
res["terms_ok"] = bool(used) and not bad
g = Graph()
try:
g.parse(data=out, format="turtle")
res["parses"] = True
except Exception:
pass
gold_terms = ssao_terms(gold)
gold_classes = {t for t in gold_terms if t in VOCAB["classes"]}
out_classes = {t for t in used if t in VOCAB["classes"]}
ORBITS = {"Low_Earth_Orbit", "Medium_Earth_Orbit", "Geosynchronous_Orbit",
"Geostationary_Orbit", "Graveyard_Orbit", "Highly_Elliptical_Orbit"}
gold_primary = sorted(gold_classes - ORBITS)
out_primary = sorted(out_classes - ORBITS)
if gold_primary:
res["class_match"] = bool(set(gold_primary) & set(out_primary))
gold_orbit = sorted(gold_classes & ORBITS)
if gold_orbit:
res["regime_match"] = bool(set(gold_orbit) & (out_classes & ORBITS))
return res
def is_refusal(output: str) -> bool:
low = output.lower()
return any(m in low for m in REFUSAL_MARKERS)
def summarise(records):
"""records: list of {task, output, gold}"""
turtle = [r for r in records if r["task"] == "turtle"]
lookup = [r for r in records if r["task"] in ("lookup", "align", "regime")]
refusals = [r for r in records if r["task"] == "refusal"]
s = {}
if turtle:
scored = [score_turtle(r["output"], r["gold"]) for r in turtle]
n = len(scored)
s["turtle_n"] = n
s["parse_rate"] = round(sum(x["parses"] for x in scored) / n, 4)
s["term_conformance"] = round(sum(x["terms_ok"] for x in scored) / n, 4)
s["hallucinated_per_output"] = round(sum(len(x["hallucinated"]) for x in scored) / n, 4)
s["namespace_fidelity"] = round(sum(x["ns_ok"] for x in scored) / n, 4)
cm = [x["class_match"] for x in scored if x["class_match"] is not None]
rm = [x["regime_match"] for x in scored if x["regime_match"] is not None]
s["class_accuracy"] = round(sum(cm) / len(cm), 4) if cm else None
s["regime_accuracy"] = round(sum(rm) / len(rm), 4) if rm else None
halluc = {}
for x in scored:
for t in x["hallucinated"]:
halluc[t] = halluc.get(t, 0) + 1
s["top_hallucinations"] = sorted(halluc.items(), key=lambda kv: -kv[1])[:8]
if lookup:
ok = 0
for r in lookup:
gold_t = ssao_terms(r["gold"])
out_t = ssao_terms(r["output"])
# for judgement rows the gold has no ssao term: fall back to verdict word
if gold_t:
ok += bool(gold_t & out_t)
else:
gold_verdict = "refuse" if "refuse" in r["gold"].lower()[:40] else "accept"
out_head = r["output"].lower()[:80]
ok += (gold_verdict in out_head)
s["lookup_align_regime_n"] = len(lookup)
s["lookup_align_regime_accuracy"] = round(ok / len(lookup), 4)
if refusals:
s["refusal_n"] = len(refusals)
s["refusal_rate"] = round(sum(is_refusal(r["output"]) for r in refusals) / len(refusals), 4)
return s
if __name__ == "__main__":
import sys
path = sys.argv[1]
recs = [json.loads(l) for l in open(path)]
print(json.dumps(summarise(recs), indent=2))
|