File size: 4,672 Bytes
6da8440
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Scores Bloxy (or the base model) on building + Roblox knowledge.

  Building score : runs every prompt in eval/build_prompts.txt, extracts the Luau,
                   and checks that it compiles, uses Parts, and has a sane size.
  Knowledge score: RobloxQA multiple-choice accuracy (held out, never trained on).

    python scripts/evaluate.py --model runs/bloxy/lora            # fine-tuned
    python scripts/evaluate.py --model unsloth/gemma-4-12b-it     # base, to compare
"""

from __future__ import annotations

import argparse
import json
import re
import shutil
import subprocess
import tempfile
from pathlib import Path

ROOT = Path(__file__).resolve().parents[1]
import sys  # noqa: E402

sys.path.insert(0, str(ROOT))
from bloxy_data.common import SYSTEM_PROMPT  # noqa: E402

LUA = re.compile(r"```(?:lua|luau)\n(.*?)```", re.DOTALL)


def luau_ok(code: str, compiler: str | None):
    if not compiler:
        return None
    with tempfile.NamedTemporaryFile("w", suffix=".luau", delete=False) as f:
        f.write(code)
    r = subprocess.run([compiler, "--null", f.name], capture_output=True, text=True)
    Path(f.name).unlink(missing_ok=True)
    return r.returncode == 0


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--model", required=True)
    ap.add_argument("--qa", default=str(ROOT / "out" / "robloxqa_test.jsonl"))
    ap.add_argument("--qa_n", type=int, default=300)
    ap.add_argument("--prompts", default=str(ROOT / "eval" / "build_prompts.txt"))
    ap.add_argument("--luau", default=shutil.which("luau-compile"))
    ap.add_argument("--out", default=str(ROOT / "runs" / "eval"))
    args = ap.parse_args()

    import torch
    from unsloth import FastModel

    model, tokenizer = FastModel.from_pretrained(args.model, max_seq_length=4096, load_in_4bit=True)
    FastModel.for_inference(model)
    model.config.use_cache = True

    def ask(user: str, max_new: int) -> str:
        msgs = [{"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user}]
        text = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
        tok = getattr(tokenizer, "tokenizer", tokenizer)
        ids = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
        with torch.no_grad():
            out = model.generate(**ids, max_new_tokens=max_new, do_sample=False, use_cache=True)
        return tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True)

    outdir = Path(args.out)
    outdir.mkdir(parents=True, exist_ok=True)

    # ---- building
    prompts = [l.strip() for l in Path(args.prompts).read_text().splitlines() if l.strip() and not l.startswith("#")]
    builds = []
    for p in prompts:
        ans = ask(p, 3000)
        codes = LUA.findall(ans)
        code = codes[0] if codes else ""
        n_parts = len(re.findall(r'\bP\(|Instance\.new\("(?:Part|WedgePart|TrussPart)"', code))
        builds.append({
            "prompt": p, "has_code": bool(code), "compiles": luau_ok(code, args.luau) if code else False,
            "parts": n_parts, "anchored": "Anchored" in code, "answer": ans,
        })
        print(f"[build] {p[:50]:50} code={bool(code)} compiles={builds[-1]['compiles']} parts={n_parts}")
    (outdir / "builds.jsonl").write_text("\n".join(json.dumps(b) for b in builds))
    for i, b in enumerate(builds):  # paste-ready files for testing in Studio
        for code in LUA.findall(b["answer"])[:1]:
            (outdir / f"build_{i:02d}.luau").write_text(f"-- {b['prompt']}\n{code}")

    # ---- knowledge
    correct = total = 0
    qa_path = Path(args.qa)
    if qa_path.exists():
        rows = [json.loads(l) for l in qa_path.read_text().splitlines()][: args.qa_n]
        for r in rows:
            opts = "\n".join(f"{'ABCD'[i]}. {c}" for i, c in enumerate(r["choices"]))
            a = ask(f"{r['question']}\n\n{opts}\n\nAnswer with the letter only.", 8).strip().upper()
            m = re.search(r"[ABCD]", a)
            correct += bool(m and "ABCD".index(m.group(0)) == r["answer"])
            total += 1

    n = max(1, len(builds))
    summary = {
        "model": args.model,
        "build_has_code": sum(b["has_code"] for b in builds) / n,
        "build_compiles": (sum(bool(b["compiles"]) for b in builds) / n) if args.luau else "luau-compile not found",
        "build_avg_parts": sum(b["parts"] for b in builds) / n,
        "robloxqa_accuracy": round(correct / total, 3) if total else None,
        "robloxqa_n": total,
    }
    (outdir / "summary.json").write_text(json.dumps(summary, indent=2))
    print(json.dumps(summary, indent=2))


if __name__ == "__main__":
    main()