File size: 6,868 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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
"""Turn correct-by-construction IES graphs into training pairs.
1. Normalise each graph's instance IRIs to clean sequential data:iN (learnable targets).
2. Emit a deterministic (description -> turtle) pair for every graph.
3. For a fraction, add a local-Qwen paraphrase of the description (linguistic diversity).
All turtle is re-validated after normalisation. Runs under .venv311."""
import sys, json, pathlib, argparse, urllib.request, re
sys.path.insert(0, str(pathlib.Path(__file__).parent))
from rdflib import Graph, URIRef, Namespace
from iesval import validate_turtle

ROOT = pathlib.Path("/Users/fabio/projects/qwen-ies-ft")
TESTDATA = "http://data.gov.uk/testdata#"
DATA = Namespace(TESTDATA)
IES = Namespace("http://ies.data.gov.uk/ontology/ies4#")
ISO = Namespace("http://iso.org/iso8601#")
URL = "http://localhost:8080/v1/chat/completions"
MODEL = "mlx-community/Qwen3.6-35B-A3B-8bit"
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.")

def _slug(v):
    s = re.sub(r"[^0-9A-Za-z]+", "_", v.strip().lower()).strip("_")
    return s[:40] or "x"

def normalise(ttl: str) -> str:
    """Rename instance IRIs to human-plausible slugs: named entities get name-based IRIs
    (data:fred_smith), unnamed nodes get type-based ones (data:boundingstate_1)."""
    g = Graph(); g.parse(data=ttl, format="turtle")
    from rdflib import RDF
    HASNAME = IES["hasName"]; REPVAL = IES["representationValue"]
    inst = sorted({str(n) for t in g for n in t
                   if isinstance(n, URIRef) and str(n).startswith(TESTDATA)})
    # collect name values per entity
    names = {}
    for s, _, o in g.triples((None, HASNAME, None)):
        v = g.value(o, REPVAL)
        if v: names.setdefault(str(s), []).append(str(v))
    remap = {}; used = {}
    def assign(u, base):
        n = used.get(base, 0) + 1; used[base] = n
        remap[URIRef(u)] = DATA[base if n == 1 else f"{base}_{n}"]
    for u in inst:
        if u in names:
            assign(u, _slug("_".join(sorted(names[u])[:2])))
    for u in inst:
        if URIRef(u) in remap: continue
        t = g.value(URIRef(u), RDF.type)
        base = _slug(str(t).split("#")[-1]) if t else "node"
        # name nodes get tied to their value where possible
        v = g.value(URIRef(u), REPVAL)
        if v and "name" in base: base = _slug(str(v)) + "_name"
        elif v: base = base + "_" + _slug(str(v))[:20]
        assign(u, base)
    ng = Graph()
    ng.bind("data", DATA); ng.bind("ies", IES); ng.bind("iso8601", ISO)
    for s, p, o in g:
        ng.add((remap.get(s, s), remap.get(p, p), remap.get(o, o)))
    return ng.serialize(format="turtle")

ALT_NS = ["http://example.org/case#", "http://ops.example.com/data#",
          "http://records.example.gov.uk/inv#", "http://intel.example.net/kg#"]

def vary_namespace(ttl, desc, i):
    """For ~35% of records swap the instance namespace and instruct the model to use it —
    teaches namespace-following instead of hardcoding testdata."""
    if i % 20 >= 7: return ttl, desc, None
    ns = ALT_NS[i % len(ALT_NS)]
    return ttl.replace(TESTDATA, ns), desc, ns

_CAPWORD = re.compile(r"\b[A-Z][a-z]+\b")
def paraphrase_ok(facts, para):
    """Every proper noun and every year in the facts must survive the paraphrase."""
    need = set(_CAPWORD.findall(facts)) | set(re.findall(r"\b(?:19|20)\d\d\b", facts))
    have = para  # substring check is enough
    return all(w in have for w in need)

def user_turn(desc, ns=None):
    ns_line = (f" Use <{ns}> as the namespace for instance IRIs." if ns else "")
    return ("Encode the following scenario as IES4 RDF/Turtle. Use only real IES4 terms and "
            f"the 4D state/period pattern where relevant.{ns_line} Output only Turtle.\n\n"
            f"Scenario: {desc}")

def rec(desc, ttl, ns=None):
    return {"messages":[{"role":"system","content":SYS},
                        {"role":"user","content":user_turn(desc, ns)},
                        {"role":"assistant","content":ttl.strip()}]}

def qwen(prompt, max_tokens=200, temperature=0.8):
    body=json.dumps({"model":MODEL,"messages":[{"role":"user","content":prompt}],
        "max_tokens":max_tokens,"temperature":temperature,
        "chat_template_kwargs":{"enable_thinking":False}}).encode()
    req=urllib.request.Request(URL, body, {"Content-Type":"application/json"})
    with urllib.request.urlopen(req, timeout=180) as r:
        return json.loads(r.read())["choices"][0]["message"].get("content","").strip()

def main():
    ap=argparse.ArgumentParser(); ap.add_argument("--frac", type=float, default=0.6)
    args=ap.parse_args()
    rows=[json.loads(l) for l in (ROOT/"data"/"ground.jsonl").open()]
    log=(ROOT/"data"/"describe.log").open("w")
    # --- pass 1: normalise + write ALL deterministic pairs (fast, no LLM) ---
    norm=[]; det=0; bad=0; varied=0
    with (ROOT/"data"/"pairs_ies.jsonl").open("w") as out:
        for i,r in enumerate(rows):
            try: ttl=normalise(r["turtle"])
            except Exception: bad+=1; continue
            if not validate_turtle(ttl,5)[0]: bad+=1; continue
            vttl, desc, ns = vary_namespace(ttl, r["facts"], i)
            if ns: varied+=1
            out.write(json.dumps(rec(desc, vttl, ns))+"\n"); det+=1
            norm.append((r["facts"], vttl, ns))
    log.write(f"pass1 deterministic={det} bad={bad} ns-varied={varied}\n"); log.flush()
    print(f"pass1: {det} deterministic pairs written (bad={bad}, ns-varied={varied})")
    # --- pass 2: append Qwen paraphrases for diversity (interruptible) ---
    para=0; rejected=0
    with (ROOT/"data"/"pairs_ies.jsonl").open("a") as out:
        for i,(facts,ttl,ns) in enumerate(norm):
            if (i % 1000)/1000.0 >= args.frac: continue
            try:
                p=qwen("Rewrite this factual statement as a natural, plainly-worded sentence or two "
                       "for a case file. Keep every fact identical; keep all names and years exactly "
                       "as written; do not add or remove facts; do not mention RDF or ontologies. "
                       f"Output only the rewrite.\n\n{facts}")
                p=p.strip().strip('"')
                if 15 < len(p) < 400 and p.lower()!=facts.lower() and paraphrase_ok(facts, p):
                    out.write(json.dumps(rec(p, ttl, ns))+"\n"); out.flush(); para+=1
                else:
                    rejected+=1
            except Exception as e:
                log.write(f"{i} qwen err {e}\n")
            if i%25==0: log.write(f"pass2 {i}/{len(norm)} para={para} rej={rejected}\n"); log.flush()
    print(f"IES pairs: deterministic={det} paraphrased={para} (rejected {rejected}) total={det+para}")

if __name__=="__main__":
    main()