File size: 8,825 Bytes
bb6a031
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
"""Evaluate baseline and trained defender on the frozen hold-out set.

Two models are compared by default:

  * **Baseline**: vanilla Qwen2.5-3B-Instruct, no SFT, no GRPO.
  * **Trained**:  Qwen2.5-3B-Instruct + SFT warm-start + GRPO curriculum.

Both are scored on `data/holdout.jsonl` using the verifier's ground-truth
labels.  Reported metrics (printed and saved to `--out-dir`):

  * Macro F1 + per-class precision/recall
  * 5x5 confusion matrix
  * Dismiss-on-malicious rate (the cardinal SOC failure mode)
  * Over-react rate (containment on benign)

Inference path
--------------
We use Unsloth's `FastLanguageModel.from_pretrained(... load_in_4bit=True)`
with `model.fast_generate` to keep eval under 10 minutes on a T4.  When
GPU deps aren't available (e.g. the Hugging Face Space build log), the
script falls back to a verifier-only sanity check by re-grading the
held-out file against itself, which serves as a smoke test.
"""

from __future__ import annotations

import argparse
import json
import os
import sys
from typing import List, Tuple

_HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.dirname(_HERE))

from eval.metrics import (  # noqa: E402
    accuracy,
    confusion_matrix,
    dismiss_on_malicious_rate,
    over_react_rate,
    per_class_f1,
)
from schema import Alert, Event, IncidentCategory, TriageAction  # noqa: E402
from train.prompt_format import (  # noqa: E402
    SYSTEM_PROMPT,
    parse_defender_response,
    render_defender_prompt,
)


def _load_holdout(path: str):
    items = []
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            items.append(json.loads(line))
    return items


def _to_alert_events(rec: dict) -> Tuple[Alert, List[Event]]:
    a = rec["alert"]
    alert = Alert(
        alert_id=a["alert_id"],
        category=IncidentCategory(a["category"]),
        severity=a["severity"],
        summary=a["summary"],
        host=a.get("host", ""),
        user=a.get("user", ""),
    )
    events = [Event(**e) for e in rec["events"]]
    return alert, events


def _print_metrics(label: str, preds: List[str], truths: List[str]) -> dict:
    cm = confusion_matrix(preds, truths)
    macro_f1, per_class = per_class_f1(cm)
    acc = accuracy(preds, truths)
    miss = dismiss_on_malicious_rate(preds, truths)
    over = over_react_rate(preds, truths)
    print(f"\n=== {label} ===")
    print(f"  accuracy:                 {acc:.3f}")
    print(f"  macro F1:                 {macro_f1:.3f}")
    print(f"  dismiss-on-malicious:     {miss:.3f}")
    print(f"  over-react on benign:     {over:.3f}")
    print("  per-class:")
    for cls, m in per_class.items():
        print(f"    {cls:<18} P={m['precision']:.2f} R={m['recall']:.2f} F1={m['f1']:.2f} (n={int(m['support'])})")
    return {
        "label": label,
        "accuracy": acc,
        "macro_f1": macro_f1,
        "dismiss_on_malicious": miss,
        "over_react_rate": over,
        "per_class": per_class,
        "confusion_matrix": cm,
    }


# ---------------------------------------------------------------------------
# Inference adapters
# ---------------------------------------------------------------------------

class _VerifierOracle:
    """A 'model' that always returns the verifier's correct answer.

    Used as a smoke test when GPU deps aren't installed; it should achieve
    100% accuracy / 0% dismiss-on-malicious by construction.
    """

    name = "verifier_oracle"

    def predict(self, alert: Alert, events: List[Event], gold: dict) -> str:
        return f"Action: {gold['ground_truth']}\nCitedLog: {gold['triggering_log_id']}\nRationale: oracle"


class _AlwaysDismissBaseline:
    """A trivial baseline that always says 'dismiss'."""

    name = "always_dismiss"

    def predict(self, alert: Alert, events: List[Event], gold: dict) -> str:
        return "Action: dismiss\nCitedLog: L1-0\nRationale: trivial baseline"


def _try_load_unsloth_model(model_name: str, adapter_path: str | None):
    """Load a model via Unsloth.  Returns None if GPU deps aren't installed."""
    try:
        from unsloth import FastLanguageModel
    except ImportError:
        return None
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=model_name,
        max_seq_length=2048,
        dtype=None,
        load_in_4bit=True,
    )
    if adapter_path and os.path.exists(adapter_path):
        model.load_adapter(adapter_path, adapter_name="default", is_trainable=False)
    FastLanguageModel.for_inference(model)
    return model, tokenizer


def _generate(model_pair, alert, events) -> str:
    model, tokenizer = model_pair
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": render_defender_prompt(alert, events)},
    ]
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    out = model.generate(**inputs, max_new_tokens=128, do_sample=False, temperature=0.0)
    text = tokenizer.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
    return text


# ---------------------------------------------------------------------------
# Main eval
# ---------------------------------------------------------------------------

def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--baseline", default="unsloth/Qwen2.5-3B-Instruct")
    parser.add_argument("--trained-adapter", default="checkpoints/defender_grpo/stage4_adversarial/adapter")
    parser.add_argument("--holdout", default="data/holdout.jsonl")
    parser.add_argument("--out-dir", default="eval/results")
    parser.add_argument("--smoke-only", action="store_true",
                        help="Skip GPU model loading; run oracle + always_dismiss only.")
    args = parser.parse_args()

    holdout_path = os.path.join(os.path.dirname(_HERE), args.holdout)
    out_dir = os.path.join(os.path.dirname(_HERE), args.out_dir)
    os.makedirs(out_dir, exist_ok=True)
    holdout = _load_holdout(holdout_path)
    truths = [r["ground_truth"] for r in holdout]
    print(f"Loaded {len(holdout)} hold-out incidents from {holdout_path}")

    summaries = []

    # --- Always-dismiss baseline (sanity) ---
    preds_dismiss = []
    for rec in holdout:
        alert, events = _to_alert_events(rec)
        text = _AlwaysDismissBaseline().predict(alert, events, rec)
        parsed = parse_defender_response(text)
        preds_dismiss.append(parsed.action.value if parsed.action else "dismiss")
    summaries.append(_print_metrics("always_dismiss", preds_dismiss, truths))

    # --- Verifier oracle (sanity) ---
    preds_oracle = []
    for rec in holdout:
        alert, events = _to_alert_events(rec)
        text = _VerifierOracle().predict(alert, events, rec)
        parsed = parse_defender_response(text)
        preds_oracle.append(parsed.action.value if parsed.action else "dismiss")
    summaries.append(_print_metrics("verifier_oracle", preds_oracle, truths))

    # --- Real models ---
    if not args.smoke_only:
        baseline_pair = _try_load_unsloth_model(args.baseline, adapter_path=None)
        if baseline_pair is not None:
            preds_baseline = []
            for rec in holdout:
                alert, events = _to_alert_events(rec)
                text = _generate(baseline_pair, alert, events)
                parsed = parse_defender_response(text)
                preds_baseline.append(parsed.action.value if parsed.action else "dismiss")
            summaries.append(_print_metrics("baseline_zero_shot", preds_baseline, truths))

            adapter_full = os.path.join(os.path.dirname(_HERE), args.trained_adapter)
            if os.path.exists(adapter_full):
                trained_pair = _try_load_unsloth_model(args.baseline, adapter_path=adapter_full)
                if trained_pair is not None:
                    preds_trained = []
                    for rec in holdout:
                        alert, events = _to_alert_events(rec)
                        text = _generate(trained_pair, alert, events)
                        parsed = parse_defender_response(text)
                        preds_trained.append(parsed.action.value if parsed.action else "dismiss")
                    summaries.append(_print_metrics("opensoc_grpo", preds_trained, truths))
            else:
                print(f"\n(skip) trained adapter not found at {adapter_full}")
        else:
            print("\n(skip) GPU deps not installed; skipping baseline_zero_shot and opensoc_grpo.")

    out_json = os.path.join(out_dir, "summary.json")
    with open(out_json, "w") as f:
        json.dump(summaries, f, indent=2)
    print(f"\nSaved summary to {out_json}")


if __name__ == "__main__":
    main()