File size: 6,990 Bytes
902cd29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
from __future__ import annotations

import json
import re
from collections import Counter, defaultdict
from dataclasses import dataclass
from pathlib import Path
from typing import Any

from db.store import Store
from visualizer.pyvis_renderer import _build_network


OUTCOME_COLORS = {
    "well_learned": "#22c55e",
    "partially_learned": "#f59e0b",
    "failed": "#ef4444",
    "not_visited": "#6b7280",
}


@dataclass(frozen=True)
class TrainingSummary:
    run_id: str
    episodes: int
    steps: int
    avg_reward: float
    avg_judge: float
    dpo_pairs: int
    top_failures: list[str]
    top_successes: list[str]


def _outcome(avg_reward: float, judge_score: float, wrong_attr_count: int, touched: bool) -> str:
    if not touched:
        return "not_visited"
    if avg_reward > 0.7 and judge_score > 0.7 and wrong_attr_count == 0:
        return "well_learned"
    if avg_reward < 0.4 or wrong_attr_count > 0:
        return "failed"
    return "partially_learned"


def build_training_graph(*, source_root: str, run_id: str, db_path: str | None = None, output_path: str = "outputs/NodeAudit_graph.html") -> Path:
    store = Store(source_root=source_root, db_path=db_path)
    snapshot = store.get_full_graph()
    annotations = store.get_training_annotations(run_id)

    by_module: dict[str, list[Any]] = defaultdict(list)
    for item in annotations:
        by_module[item.module_id].append(item)

    net = _build_network(height="920px", width="100%")

    failed_edges: set[tuple[str, str]] = set()
    all_rewards: list[float] = []
    all_judges: list[float] = []

    for node in snapshot.nodes:
        rows = by_module.get(node.module_id, [])
        touched = bool(rows)
        rewards = [float(row.avg_reward) for row in rows]
        judges = [float(row.thinking_quality) for row in rows]
        avg_reward = (sum(rewards) / len(rewards)) if rewards else 0.0
        avg_judge = (sum(judges) / len(judges)) if judges else 0.0
        all_rewards.extend(rewards)
        all_judges.extend(judges)

        action_counts: Counter[str] = Counter()
        correct: list[str] = []
        wrong: list[str] = []
        judge_text: list[str] = []

        for row in rows:
            try:
                action_counts.update(json.loads(row.action_counts_json))
            except Exception:
                if row.action_type:
                    action_counts[row.action_type] += 1
            try:
                correct.extend(json.loads(row.correct_attributions_json))
            except Exception:
                pass
            try:
                wrong.extend(json.loads(row.wrong_attributions_json))
            except Exception:
                pass
            if row.judge_verdict:
                judge_text.append(row.judge_verdict)

            if row.action_type == "FLAG_DEPENDENCY_ISSUE":
                try:
                    payload = json.loads(row.action_payload)
                except Exception:
                    payload = {}
                target = str(payload.get("attributed_to") or "")
                if target and wrong:
                    failed_edges.add((node.module_id, target))

        outcome = _outcome(avg_reward, avg_judge, len(wrong), touched)
        actions_pretty = ", ".join(f"{k}x{v}" for k, v in sorted(action_counts.items())) or "none"
        judge_verdict = judge_text[-1] if judge_text else "not judged"

        tooltip = (
            f"Module: {node.module_id}\n"
            f"Avg Reward: {avg_reward:.2f}\n"
            f"Judge Score: {avg_judge:.2f}\n"
            f"Correct Attributions: {', '.join(correct) if correct else 'none'}\n"
            f"Wrong: {', '.join(wrong) if wrong else 'none'}\n"
            f"Actions: {actions_pretty}\n"
            f"Judge Verdict: {judge_verdict}"
        )

        net.add_node(
            n_id=node.module_id,
            label=node.module_id,
            title=tooltip,
            color=OUTCOME_COLORS[outcome],
            value=1.0 + max(0.0, avg_reward),
            shape="dot",
        )

    for edge in snapshot.edges:
        is_failed = (edge.source_module_id, edge.target_module_id) in failed_edges
        net.add_edge(
            source=edge.source_module_id,
            to=edge.target_module_id,
            title=edge.connection_summary or edge.import_line,
            color="#ef4444" if is_failed else "#2563eb",
            width=2.2 if is_failed else 1.4,
            arrows="to",
        )

    summary = _summarize(run_id=run_id, annotations=annotations, rewards=all_rewards, judges=all_judges)

    output = Path(output_path).resolve()
    output.parent.mkdir(parents=True, exist_ok=True)
    net.write_html(str(output), open_browser=False, notebook=False)

    html = output.read_text(encoding="utf-8")
    html = re.sub(r'<link[^>]*cdn\.jsdelivr\.net[^>]*>\s*', "", html, flags=re.IGNORECASE)
    html = re.sub(r'<script[^>]*cdn\.jsdelivr\.net[^>]*>\s*</script>\s*', "", html, flags=re.IGNORECASE)
    panel = (
        "<aside style='position:fixed;right:0;top:0;width:340px;height:100%;"
        "background:#0f172a;color:#e2e8f0;padding:18px;overflow:auto;z-index:1000;'>"
        f"<h3 style='margin:0 0 12px 0;'>Training Run: {summary.run_id}</h3>"
        f"<p>Episodes: {summary.episodes} | Steps: {summary.steps}</p>"
        f"<p>Avg Reward: {summary.avg_reward:.2f}</p>"
        f"<p>Judge Scores: {summary.avg_judge:.2f}</p>"
        f"<p>DPO pairs built: {summary.dpo_pairs}</p>"
        "<h4>Top 3 Failures</h4>"
        f"<ul>{''.join(f'<li>{item}</li>' for item in summary.top_failures)}</ul>"
        "<h4>Top 3 Successes</h4>"
        f"<ul>{''.join(f'<li>{item}</li>' for item in summary.top_successes)}</ul>"
        "</aside>"
    )

    html = html.replace("</body>", f"{panel}</body>")
    output.write_text(html, encoding="utf-8")
    return output


def _summarize(*, run_id: str, annotations: list[Any], rewards: list[float], judges: list[float]) -> TrainingSummary:
    episodes = len({(item.task_id, item.module_id) for item in annotations})
    steps = len(annotations)
    avg_reward = (sum(rewards) / len(rewards)) if rewards else 0.0
    avg_judge = (sum(judges) / len(judges)) if judges else 0.0

    module_scores: dict[str, float] = defaultdict(float)
    module_counts: dict[str, int] = defaultdict(int)
    for row in annotations:
        module_scores[row.module_id] += float(row.avg_reward)
        module_counts[row.module_id] += 1

    sorted_modules = sorted(
        module_scores,
        key=lambda module_id: module_scores[module_id] / max(1, module_counts[module_id]),
    )
    top_failures = sorted_modules[:3]
    top_successes = list(reversed(sorted_modules[-3:])) if sorted_modules else []

    return TrainingSummary(
        run_id=run_id,
        episodes=episodes,
        steps=steps,
        avg_reward=avg_reward,
        avg_judge=avg_judge,
        dpo_pairs=max(0, steps // 4),
        top_failures=top_failures,
        top_successes=top_successes,
    )