File size: 6,190 Bytes
e8c001c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
#!/usr/bin/env python3
"""
flinksql_005: End-to-end verification that ground_truth.sql earns FULL product score.

GT is graded as if it were the agent's answer. We assert the ground truth earns
the FULL product score (A/B/C/D dims, product_ratio == 1.0). Process dims
(G/H/I/J) are ignored: a bare GT run has no agent transcript, so those dims are
legitimately 0 and must NOT count against the case.

Flow:
  1. Start Docker container with Flink environment
  2. Setup workspace (copy init/)
  3. Copy ground_truth.sql -> result.sql
  4. Inject gt/ directory
  5. Run grade(workspace_path='/tmp_workspace')
  6. Assert product_ratio == 1.0 (GT must earn full product score)
  7. Cleanup container

Usage:
    python3 verify_grade.py
"""
import json
import os
import subprocess
import sys
import tempfile
import uuid
from pathlib import Path

DOCKER_IMAGE = os.environ.get("DOCKER_IMAGE", "dataclaw-eval:v1.0")
TMP_WORKSPACE = "/tmp_workspace"
TASK_DIR = Path(__file__).resolve().parent.parent  # flinksql_005/
CONTAINER_NAME = f"verify_gt_flinksql_005_{uuid.uuid4().hex[:8]}"

PRODUCT_PREFIXES = ("A_", "B_", "C_", "D_")


def run_cmd(cmd, timeout=300, check=True):
    """Run a command, raise on failure if check=True."""
    r = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
                       universal_newlines=True, timeout=timeout)
    if check and r.returncode != 0:
        raise RuntimeError("Command failed: {}\nstderr: {}\nstdout: {}".format(
            ' '.join(cmd), r.stderr, r.stdout))
    return r


def cleanup():
    subprocess.run(["docker", "rm", "-f", CONTAINER_NAME],
                   stdout=subprocess.PIPE, stderr=subprocess.PIPE)


def product_ratio(scores):
    """GT product ratio (0..1) = sum(A/B/C/D raw score) / sum(A/B/C/D max).

    Uses RAW dimension scores, NOT product_points: product_points already
    applies a per-case product_weight (0.6 or 0.7) which differs across cases
    and is irrelevant to "did GT max out the product dims".
    """
    raw = mx = 0.0
    for dim, info in (scores.get("details", {}) or {}).items():
        if dim.startswith(PRODUCT_PREFIXES) and isinstance(info, dict):
            raw += info.get("score", 0) or 0
            mx += info.get("max", 0) or 0
    return round(raw / mx, 4) if mx else 0.0


def main():
    workspace_path = str(TASK_DIR)
    print(f"[verify] Task dir: {workspace_path}")
    print(f"[verify] Container: {CONTAINER_NAME}")
    print(f"[verify] Docker image: {DOCKER_IMAGE}")

    cleanup()

    try:
        print("[verify] Step 1: Starting container...")
        run_cmd([
            "docker", "run", "-d",
            "--name", CONTAINER_NAME,
            "-v", f"{workspace_path}/init:/app:ro",
            DOCKER_IMAGE, "/bin/bash", "-c", "tail -f /dev/null",
        ])

        print("[verify] Step 2: Setting up workspace...")
        run_cmd([
            "docker", "exec", CONTAINER_NAME, "/bin/bash", "-c",
            f"mkdir -p {TMP_WORKSPACE}/init && cp -r /app/. {TMP_WORKSPACE}/init/ "
            f"&& chmod -R u+w {TMP_WORKSPACE}",
        ])

        print("[verify] Step 3: Copying ground_truth.sql -> result.sql...")
        gt_host = os.path.join(workspace_path, "gt")
        run_cmd(["docker", "cp", gt_host, f"{CONTAINER_NAME}:{TMP_WORKSPACE}/gt"])
        run_cmd([
            "docker", "exec", CONTAINER_NAME, "/bin/bash", "-c",
            f"cp {TMP_WORKSPACE}/gt/ground_truth.sql {TMP_WORKSPACE}/result.sql",
        ])

        print("[verify] Step 4: Running grade()...")
        grade_runner = "\n".join([
            "import json",
            "import sys",
            f"sys.path.insert(0, '{TMP_WORKSPACE}')",
            f"sys.path.insert(0, '{TMP_WORKSPACE}/gt')",
            "from grade import grade",
            f"result = grade(workspace_path='{TMP_WORKSPACE}')",
            "print('__VERIFY_JSON_START__')",
            "print(json.dumps(result, ensure_ascii=False))",
            "print('__VERIFY_JSON_END__')",
        ]) + "\n"

        with tempfile.NamedTemporaryFile(mode="w", suffix=".py", delete=False,
                                         encoding="utf-8") as f:
            f.write(grade_runner)
            runner_host = f.name

        try:
            run_cmd(["docker", "cp", runner_host,
                     f"{CONTAINER_NAME}:/tmp/_verify_runner.py"])
            r = subprocess.run(
                ["docker", "exec", CONTAINER_NAME, "python3", "/tmp/_verify_runner.py"],
                stdout=subprocess.PIPE, stderr=subprocess.PIPE,
                universal_newlines=True, timeout=600,
            )
        finally:
            os.unlink(runner_host)

        stdout = r.stdout or ""
        if "__VERIFY_JSON_START__" not in stdout or "__VERIFY_JSON_END__" not in stdout:
            print("[verify] FAIL: No JSON output from grade")
            print(f"  stdout (last 2000): {stdout[-2000:]}")
            print(f"  stderr (last 2000): {(r.stderr or '')[-2000:]}")
            sys.exit(1)

        json_str = stdout.split("__VERIFY_JSON_START__")[1].split("__VERIFY_JSON_END__")[0].strip()
        scores = json.loads(json_str)

        print("[verify] Grade result:")
        print(json.dumps(scores, indent=2, ensure_ascii=False))

        ratio = product_ratio(scores)
        print(f"[verify] product_ratio = {ratio} "
              f"(product_points={scores.get('product_points')}, "
              f"overall_score={scores.get('overall_score')})")

        if ratio >= 1.0:
            print(f"\n[verify] PASS: GT earned full product score (product_ratio={ratio})")
        else:
            print(f"\n[verify] FAIL: GT did NOT earn full product score (product_ratio={ratio})")
            for dim, info in (scores.get("details", {}) or {}).items():
                if dim.startswith(PRODUCT_PREFIXES) and isinstance(info, dict):
                    print(f"  {dim}: {info.get('score', 0)}/{info.get('max', 0)}")
            for d in (scores.get("diagnostics") or []):
                print(f"  diag: {d}")
            sys.exit(1)

    except Exception as e:
        print(f"[verify] ERROR: {e}")
        sys.exit(1)
    finally:
        cleanup()


if __name__ == "__main__":
    main()