File size: 27,145 Bytes
030876e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
"""
eval_gpt5_results.py
---------------------
Evaluate pre-generated GPT-5 inference results (from run_gpt5_inference.py)
with the same metrics used by test_classifier_with_subclaim_thresholds.py:

  1. Classifier accuracy  (DSPy health-literacy classifier)
  2. Completeness score   (recall: summary_subclaims covered by gen_text)
  3. Hallucination score  (gen_text sentences NOT supported by input_text)

Expected JSONL format (from run_gpt5_inference.py): each line has model,
row_index, doc_id, gold_label, source_lang, prompt, prediction, generated_text,
error. Reference (--reference-file) supplies summary_subclaims and input_text
by (doc_id, gold_label).

Usage
-----
# Offline: count scores only (no classifier/support API required)
python eval_gpt5_results.py --input-file gpt5mini-nano_inference/gpt5_inference_gpt-5_20260302_201653.jsonl --offline

# Full evaluation (requires classifier API + support API + dspy)
python eval_gpt5_results.py --input-file gpt5mini-nano_inference/gpt5_inference_gpt-5_20260302_201653.jsonl

# Multiple files
python eval_gpt5_results.py --input-file file1.jsonl file2.jsonl
"""

import argparse
import json
import os
import re
import traceback
import urllib.error
import urllib.request
from datetime import datetime
from typing import Any, Dict, List, Optional, Tuple

try:
    import dspy
except ImportError:
    dspy = None  # type: ignore[assignment]
import requests
from tqdm import tqdm


# ---------------------------------------------------------------------------
# Defaults
# ---------------------------------------------------------------------------

DEFAULT_CLASSIFIER_API_BASE = "http://172.16.34.19:8040/v1"
DEFAULT_SUPPORT_API_BASE    = "http://172.16.34.19:8090"
DEFAULT_MODEL_PATH = (
    "/home/mshahidul/readctrl/code/readctrl_rl_inference/model.json"
)
DEFAULT_REFERENCE_FILE = (
    "/home/mshahidul/readctrl/code/text_classifier/data/"
    "verified_combined_0-80_clean200_with_subclaims.json"
)
DEFAULT_OUTPUT_DIR = (
    "/home/mshahidul/readctrl/code/readctrl_rl_inference/test_result_v4"
)

VALID_LABELS = {
    "low_health_literacy",
    "intermediate_health_literacy",
    "proficient_health_literacy",
}

MIN_SENTENCE_CHARS = 15


# ---------------------------------------------------------------------------
# Sentence splitter (mirrors reward_new_v5.py)
# ---------------------------------------------------------------------------

def _split_into_sentences(text: str, min_chars: int = MIN_SENTENCE_CHARS) -> List[str]:
    if not text or not text.strip():
        return []
    parts = re.split(r"(?<=[.!?])\s+", text.strip())
    return [s.strip() for s in parts if len(s.strip()) >= min_chars]


# ---------------------------------------------------------------------------
# DSPy health-literacy classifier (only when dspy is available)
# ---------------------------------------------------------------------------

if dspy is not None:
    class HealthLiteracySignature(dspy.Signature):
        generated_text = dspy.InputField(
            desc="A version of the source text rewritten for a specific audience."
        )
        literacy_label = dspy.OutputField(
            desc=(
                "Classification: low_health_literacy (simple words, no jargon), "
                "intermediate_health_literacy (moderate technicality), or "
                "proficient_health_literacy (highly technical/original level)."
            )
        )

    class HealthLiteracyClassifier(dspy.Module):
        def __init__(self):
            super().__init__()
            self.classifier = dspy.ChainOfThought(HealthLiteracySignature)

        def forward(self, generated_text):
            return self.classifier(generated_text=generated_text)
else:
    HealthLiteracyClassifier = None  # type: ignore[misc, assignment]


# ---------------------------------------------------------------------------
# Support-API verifier (mirrors reward_new_v5.py + test_classifier script)
# ---------------------------------------------------------------------------

class MedicalClaimVerifier:
    """
    Calls FastAPI POST /check_support.
    base_url: 'http://host:8090'  β€” NO /v1 suffix.
    """

    def __init__(self, base_url: str):
        self.base_url = base_url.rstrip("/")

    def _call_support_api(
        self,
        context: str,
        subclaims: List[str],
        threshold: float = 0.5,
        batch_size: int = 128,
    ) -> Optional[List[str]]:
        """Returns label list or None on total network failure."""
        if not context or not subclaims:
            return ["invalid"] * len(subclaims)
        try:
            resp = requests.post(
                f"{self.base_url}/check_support",
                json={
                    "context": context,
                    "subclaims": subclaims,
                    "threshold": threshold,
                    "batch_size": batch_size,
                },
                timeout=300,
            )
            resp.raise_for_status()
            labels = resp.json().get("labels", ["invalid"] * len(subclaims))
            if len(labels) < len(subclaims):
                labels.extend(["invalid"] * (len(subclaims) - len(labels)))
            elif len(labels) > len(subclaims):
                labels = labels[: len(subclaims)]
            return labels
        except requests.exceptions.RequestException as exc:
            print(f"Warning: Support API call failed (returning None): {exc}")
            return None

    def compute_completeness(
        self, summary_subclaims: List[str], gen_text: str
    ) -> Optional[float]:
        """Fraction of summary_subclaims covered by gen_text (recall direction)."""
        if not summary_subclaims or not gen_text or not gen_text.strip():
            return 0.0
        labels = self._call_support_api(context=gen_text, subclaims=summary_subclaims)
        if labels is None:
            return None
        valid = [l for l in labels if str(l).strip().lower() != "invalid"]
        if not valid:
            return None
        covered = sum(1 for l in valid if str(l).strip().lower() == "supported")
        return covered / len(valid)

    def compute_hallucination(
        self, input_text: str, gen_text: str
    ) -> Optional[float]:
        """Fraction of gen_text sentences NOT supported by input_text."""
        gen_segs = _split_into_sentences(gen_text)
        if not gen_segs or not input_text or not input_text.strip():
            return 0.0
        input_sents = _split_into_sentences(input_text)
        stable_denom = max(len(gen_segs), len(input_sents))
        if stable_denom == 0:
            return 0.0
        labels = self._call_support_api(context=input_text, subclaims=gen_segs)
        if labels is None:
            return None
        valid = [l for l in labels if str(l).strip().lower() != "invalid"]
        if not valid:
            return None
        hallucinated = sum(1 for l in valid if str(l).strip().lower() != "supported")
        return hallucinated / stable_denom

    def evaluate_sample(
        self, gen_text: str, summary_subclaims: List[str], input_text: str
    ) -> Tuple[Optional[float], Optional[float]]:
        completeness = self.compute_completeness(summary_subclaims, gen_text)
        hallucination = self.compute_hallucination(input_text, gen_text)
        return completeness, hallucination


# ---------------------------------------------------------------------------
# Health checks
# ---------------------------------------------------------------------------

def check_api_base(api_base: str) -> None:
    models_url = api_base.rstrip("/") + "/models"
    req = urllib.request.Request(models_url, method="GET")
    try:
        with urllib.request.urlopen(req, timeout=5) as resp:
            if resp.status >= 400:
                raise RuntimeError(f"Unhealthy endpoint: {models_url}")
    except urllib.error.URLError as exc:
        raise ConnectionError(
            f"Cannot reach classifier API: {api_base}. Start vLLM server."
        ) from exc


def check_support_api_base(api_base: str) -> None:
    url = api_base.rstrip("/") + "/check_support"
    try:
        resp = requests.post(
            url,
            json={"context": "test", "subclaims": ["test"], "threshold": 0.5, "batch_size": 1},
            timeout=5,
        )
        if resp.status_code >= 500:
            raise RuntimeError(f"Support API server error: {url}")
    except requests.exceptions.ConnectionError as exc:
        raise ConnectionError(f"Cannot reach Support API: {url}.") from exc
    except requests.exceptions.Timeout as exc:
        raise ConnectionError(f"Support API timed out: {url}") from exc


# ---------------------------------------------------------------------------
# Data loading
# ---------------------------------------------------------------------------

def load_compiled_classifier(path: str):
    if hasattr(dspy, "load"):
        try:
            return dspy.load(path)
        except Exception:
            pass
    classifier = HealthLiteracyClassifier()
    try:
        classifier.load(path)
    except Exception as exc:
        raise RuntimeError(f"Failed to load model: {path}") from exc
    return classifier


def normalize_pred_label(pred_obj: Any) -> str:
    if not pred_obj or not hasattr(pred_obj, "literacy_label"):
        return ""
    return str(pred_obj.literacy_label).strip().lower()


def load_inference_jsonl(path: str) -> List[Dict[str, Any]]:
    """
    Load GPT-5 inference JSONL produced by run_gpt5_inference.py (or
    run_gpt5mini_nano_inference.py). Expected fields per row: model,
    row_index, doc_id, gold_label, generated_text, error; optional:
    source_lang, prompt, prediction, input_text.
    Rows with non-empty 'error' or empty 'generated_text' are kept but
    flagged so they can be skipped cleanly.
    """
    items = []
    with open(path, "r", encoding="utf-8") as f:
        for line_no, line in enumerate(f, start=1):
            if not line.strip():
                continue
            row = json.loads(line)
            items.append({
                "line_no":        line_no,
                "model":          str(row.get("model", "")).strip(),
                "row_index":      row.get("row_index"),
                "doc_id":         row.get("doc_id"),
                "gold_label":     str(row.get("gold_label", "")).strip(),
                "generated_text": str(row.get("generated_text", "")).strip(),
                "input_text":     str(row.get("input_text", "")).strip(),
                "error":          str(row.get("error", "")).strip(),
            })
    return items


def load_reference_lookup(
    reference_path: str,
) -> Dict[Tuple[Any, str], Dict[str, Any]]:
    """
    Returns (doc_id, label) β†’ {summary_subclaims, input_text}.
    Falls back to 'fulltext' field for input_text if 'input_text' absent.
    """
    with open(reference_path, "r", encoding="utf-8") as f:
        rows = json.load(f)
    if not isinstance(rows, list):
        raise ValueError("Reference file must be a JSON list.")

    lookup: Dict[Tuple[Any, str], Dict[str, Any]] = {}
    for row in rows:
        doc_id = row.get("doc_id")
        label  = str(row.get("label", "")).strip()
        if label not in VALID_LABELS:
            continue
        summary_subclaims = row.get("summary_subclaims", row.get("gold_subclaims", []))
        input_text = str(row.get("input_text", row.get("fulltext", ""))).strip()
        if not isinstance(summary_subclaims, list) or not summary_subclaims:
            continue
        entry = {"summary_subclaims": summary_subclaims, "input_text": input_text}
        for key in [(doc_id, label), (str(doc_id), label)]:
            lookup.setdefault(key, entry)
    if not lookup:
        raise ValueError(f"Reference lookup is empty: {reference_path}")
    return lookup


# ---------------------------------------------------------------------------
# Offline evaluation (no classifier/support API)
# ---------------------------------------------------------------------------

def evaluate_file_offline(
    *,
    input_path: str,
    reference_lookup: Dict,
    output_dir: str,
    max_samples: int,
) -> Dict[str, Any]:
    """
    Compute basic counts and scores from inference JSONL without calling
    classifier or support API. Use --offline when those services are unavailable.
    """
    rows = load_inference_jsonl(input_path)
    model_name = next((r["model"] for r in rows if r["model"]), os.path.basename(input_path))

    if max_samples > 0:
        rows = rows[:max_samples]

    total_in_file = len(rows)
    error_rows = 0
    no_text_rows = 0
    unmatched_rows = 0
    evaluated_count = 0

    for row in rows:
        if row["error"]:
            error_rows += 1
            continue
        if not row["generated_text"]:
            no_text_rows += 1
            continue
        gold_label = row["gold_label"]
        if gold_label not in VALID_LABELS:
            continue
        doc_id = row["doc_id"]
        ref = reference_lookup.get((doc_id, gold_label)) or reference_lookup.get((str(doc_id), gold_label))
        if not ref:
            unmatched_rows += 1
            continue
        evaluated_count += 1

    score_summary = {
        "model": model_name,
        "input_file": input_path,
        "total_rows_in_file": total_in_file,
        "error_rows_skipped": error_rows,
        "rows_without_generated_text": no_text_rows,
        "unmatched_rows": unmatched_rows,
        "evaluable_rows": evaluated_count,
        "success_rate": evaluated_count / total_in_file if total_in_file else 0.0,
    }

    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_slug = model_name.replace("/", "_").replace(" ", "_")
    os.makedirs(output_dir, exist_ok=True)
    summary_path = os.path.join(output_dir, f"gpt5_eval_offline_{model_slug}_{ts}.json")
    with open(summary_path, "w", encoding="utf-8") as f:
        json.dump(score_summary, f, indent=2)
    print(json.dumps(score_summary, indent=2))
    print(f"[DONE] {model_name} (offline): summary β†’ {summary_path}")
    return score_summary


# ---------------------------------------------------------------------------
# Per-file evaluation
# ---------------------------------------------------------------------------

def evaluate_file(
    *,
    input_path: str,
    reference_lookup: Dict,
    classifier,
    verifier: MedicalClaimVerifier,
    comp_threshold: float,
    halluc_threshold: float,
    output_dir: str,
    max_samples: int,
    provide_traceback: bool,
) -> Dict[str, Any]:
    """Run evaluation on one JSONL file; save summary + details; return summary dict."""

    rows = load_inference_jsonl(input_path)
    # Detect model name from first valid row
    model_name = next((r["model"] for r in rows if r["model"]), os.path.basename(input_path))

    if max_samples > 0:
        rows = rows[:max_samples]

    # ── counters ────────────────────────────────────────────────────────────
    unmatched_rows       = 0
    error_rows           = 0
    total                = 0
    classifier_correct   = 0
    comp_pass_count      = 0
    halluc_fail_count    = 0
    cls_and_comp_count   = 0
    cls_comp_nh_count    = 0
    comp_sum, comp_n     = 0.0, 0
    halluc_sum, halluc_n = 0.0, 0
    details: List[Dict[str, Any]] = []

    CHECKPOINT_EVERY = 10
    ts = datetime.now().strftime("%Y%m%d_%H%M%S")
    model_slug = model_name.replace("/", "_").replace(" ", "_")
    os.makedirs(output_dir, exist_ok=True)
    summary_path = os.path.join(output_dir, f"gpt5_eval_{model_slug}_{ts}.json")
    details_path = os.path.join(output_dir, f"gpt5_eval_{model_slug}_{ts}.jsonl")

    def build_summary() -> Dict[str, Any]:
        safe = lambda n: n / total if total else 0.0
        return {
            "model":                          model_name,
            "input_file":                     input_path,
            "total_rows_in_file":             len(rows),
            "total_samples_evaluated":        total,
            "unmatched_rows":                 unmatched_rows,
            "error_rows_skipped":             error_rows,
            # classifier
            "classifier_only_accuracy":       safe(classifier_correct),
            # completeness
            "completeness_pass_rate":         safe(comp_pass_count),
            "completeness_mean":              comp_sum / comp_n if comp_n else None,
            "completeness_threshold":         comp_threshold,
            # hallucination
            "hallucination_fail_rate":        safe(halluc_fail_count),
            "hallucination_mean":             halluc_sum / halluc_n if halluc_n else None,
            "hallucination_threshold":        halluc_threshold,
            # combined
            "accuracy_cls_and_completeness":          safe(cls_and_comp_count),
            "accuracy_cls_comp_no_hallucination":     safe(cls_comp_nh_count),
            "details_path":                   details_path,
        }

    def save_checkpoint() -> None:
        with open(summary_path, "w", encoding="utf-8") as f:
            json.dump(build_summary(), f, indent=2)
        with open(details_path, "w", encoding="utf-8") as f:
            for item in details:
                f.write(json.dumps(item, ensure_ascii=False) + "\n")

    for idx, row in enumerate(tqdm(rows, desc=model_name), start=1):
        gold_label     = row["gold_label"]
        generated_text = row["generated_text"]
        doc_id         = row["doc_id"]

        if gold_label not in VALID_LABELS:
            continue
        if row["error"]:
            error_rows += 1
            continue
        if not generated_text:
            continue

        ref = reference_lookup.get((doc_id, gold_label)) or \
              reference_lookup.get((str(doc_id), gold_label))
        if not ref:
            unmatched_rows += 1
            continue

        summary_subclaims = ref["summary_subclaims"]
        input_text = ref.get("input_text") or row.get("input_text", "")

        total += 1

        # 1. Classifier
        pred       = classifier(generated_text=generated_text)
        pred_label = normalize_pred_label(pred)
        is_cls_correct = gold_label in pred_label
        classifier_correct += int(is_cls_correct)

        # 2. Completeness + Hallucination
        comp_score, halluc_score = verifier.evaluate_sample(
            gen_text=generated_text,
            summary_subclaims=summary_subclaims,
            input_text=input_text,
        )

        comp_pass   = (comp_score is not None) and (comp_score >= comp_threshold)
        halluc_fail = (halluc_score is not None) and (halluc_score > halluc_threshold)
        comp_pass_count   += int(comp_pass)
        halluc_fail_count += int(halluc_fail)
        if comp_score is not None:
            comp_sum += comp_score;  comp_n += 1
        if halluc_score is not None:
            halluc_sum += halluc_score;  halluc_n += 1

        cls_and_comp   = is_cls_correct and comp_pass
        cls_comp_no_h  = cls_and_comp and not halluc_fail
        cls_and_comp_count += int(cls_and_comp)
        cls_comp_nh_count  += int(cls_comp_no_h)

        details.append({
            "idx":                         idx,
            "model":                       model_name,
            "line_no":                     row.get("line_no"),
            "row_index":                   row.get("row_index"),
            "doc_id":                      doc_id,
            "gold_label":                  gold_label,
            "generated_text":              generated_text,
            "pred_label":                  pred_label,
            "classifier_correct":          is_cls_correct,
            "completeness_score":          comp_score,
            "completeness_pass":           comp_pass,
            "completeness_threshold":      comp_threshold,
            "hallucination_score":         halluc_score,
            "hallucination_fail":          halluc_fail,
            "hallucination_threshold":     halluc_threshold,
            "pass_cls_and_completeness":   cls_and_comp,
            "pass_cls_comp_no_hallucination": cls_comp_no_h,
        })

        if total % CHECKPOINT_EVERY == 0:
            save_checkpoint()
            comp_avg   = f"{comp_sum/comp_n:.4f}"   if comp_n   else "N/A"
            halluc_avg = f"{halluc_sum/halluc_n:.4f}" if halluc_n else "N/A"
            print(
                f"\n[CHECKPOINT {model_name}] {total} samples β€” "
                f"cls_acc={classifier_correct/total:.4f}, "
                f"comp_pass={comp_pass_count/total:.4f} (mean={comp_avg}), "
                f"halluc_fail={halluc_fail_count/total:.4f} (mean={halluc_avg})"
            )

    if total == 0:
        raise RuntimeError(f"No valid rows found in {input_path}")

    save_checkpoint()
    summary = build_summary()
    print(json.dumps(summary, indent=2))
    print(f"[DONE] {model_name}: summary β†’ {summary_path}")
    print(f"[DONE] {model_name}: details β†’ {details_path}")
    return summary


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------

def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(
        description=(
            "Evaluate GPT-5 mini/nano inference results: classifier accuracy, "
            "completeness (recall), and hallucination score."
        )
    )
    parser.add_argument(
        "--input-file",
        nargs="+",
        required=True,
        help=(
            "One or more JSONL files produced by run_gpt5mini_nano_inference.py. "
            "Each file is evaluated separately."
        ),
    )
    parser.add_argument("--model-path", default=DEFAULT_MODEL_PATH,
                        help="DSPy health-literacy classifier model.json path.")
    parser.add_argument("--reference-file", default=DEFAULT_REFERENCE_FILE,
                        help="Reference JSON with summary_subclaims + input_text.")
    parser.add_argument("--classifier-api-base", default=DEFAULT_CLASSIFIER_API_BASE)
    parser.add_argument(
        "--support-api-base", default=DEFAULT_SUPPORT_API_BASE,
        help="FastAPI /check_support base URL (NO /v1 suffix).",
    )
    parser.add_argument("--output-dir", default=DEFAULT_OUTPUT_DIR)
    parser.add_argument("--comp-threshold", type=float, default=0.5,
                        help="Completeness pass threshold (score >= value).")
    parser.add_argument("--hallucination-threshold", type=float, default=0.1,
                        help="Hallucination fail threshold (score > value).")
    parser.add_argument("--max-samples", type=int, default=-1,
                        help="Max rows per file. -1 = all.")
    parser.add_argument("--provide-traceback", action="store_true")
    parser.add_argument("--offline", action="store_true",
                        help="Only compute counts/success rate; no classifier or support API.")
    return parser.parse_args()


def main() -> None:
    args = parse_args()

    if not os.path.exists(args.reference_file):
        raise FileNotFoundError(f"Reference file not found: {args.reference_file}")
    for f in args.input_file:
        if not os.path.exists(f):
            raise FileNotFoundError(f"Input file not found: {f}")

    ref_lookup = load_reference_lookup(args.reference_file)

    if args.offline:
        all_summaries = []
        for input_path in args.input_file:
            print(f"\n{'='*60}")
            print(f" Evaluating (offline): {os.path.basename(input_path)}")
            print(f"{'='*60}")
            summary = evaluate_file_offline(
                input_path=input_path,
                reference_lookup=ref_lookup,
                output_dir=args.output_dir,
                max_samples=args.max_samples,
            )
            all_summaries.append(summary)
        if len(all_summaries) > 1:
            print(f"\n{'='*60}")
            print(" OFFLINE SUMMARY")
            print(f"{'='*60}")
            for s in all_summaries:
                print(f"  {s['model']}: {s['evaluable_rows']}/{s['total_rows_in_file']} evaluable, success_rate={s['success_rate']:.4f}")
        return

    if not os.path.exists(args.model_path):
        raise FileNotFoundError(f"Model file not found: {args.model_path}")
    if dspy is None:
        raise RuntimeError(
            "Full evaluation requires dspy. Install with: pip install dspy-ai"
        )

    try:
        check_api_base(args.classifier_api_base)
        check_support_api_base(args.support_api_base)

        lm = dspy.LM(
            model="openai/dspy",
            api_base=args.classifier_api_base,
            api_key="EMPTY",
            temperature=0.0,
        )
        dspy.configure(lm=lm)
        classifier   = load_compiled_classifier(args.model_path)
        verifier     = MedicalClaimVerifier(base_url=args.support_api_base)

        all_summaries: List[Dict[str, Any]] = []
        for input_path in args.input_file:
            print(f"\n{'='*60}")
            print(f" Evaluating: {os.path.basename(input_path)}")
            print(f"{'='*60}")
            summary = evaluate_file(
                input_path=input_path,
                reference_lookup=ref_lookup,
                classifier=classifier,
                verifier=verifier,
                comp_threshold=args.comp_threshold,
                halluc_threshold=args.hallucination_threshold,
                output_dir=args.output_dir,
                max_samples=args.max_samples,
                provide_traceback=args.provide_traceback,
            )
            all_summaries.append(summary)

        # ── Cross-model comparison table ────────────────────────────────────
        if len(all_summaries) > 1:
            print(f"\n{'='*60}")
            print(" CROSS-MODEL COMPARISON")
            print(f"{'='*60}")
            fmt = "{:<20} {:>10} {:>12} {:>12} {:>12} {:>14}"
            print(fmt.format(
                "Model", "CLS Acc", "Comp Pass%",
                "Comp Mean", "Halluc Fail%", "Cls+Comp+NoH%"
            ))
            print("-" * 82)
            for s in all_summaries:
                name      = s["model"][-20:]
                cls_acc   = f"{s['classifier_only_accuracy']*100:.1f}%"
                comp_pass = f"{s['completeness_pass_rate']*100:.1f}%"
                comp_mean_val = s.get("completeness_mean")
                comp_mean = f"{comp_mean_val:.4f}" if comp_mean_val is not None else "N/A"
                halluc_f  = f"{s['hallucination_fail_rate']*100:.1f}%"
                combined  = f"{s['accuracy_cls_comp_no_hallucination']*100:.1f}%"
                print(fmt.format(name, cls_acc, comp_pass, comp_mean, halluc_f, combined))

    except Exception as exc:
        print(f"[error] {type(exc).__name__}: {exc}")
        if args.provide_traceback:
            traceback.print_exc()
        raise


if __name__ == "__main__":
    main()