1use crate::{
2 PredictArgs, PredictionProvider,
3 example::{ActualCursor, Example, ExampleScore},
4 format_prompt::TeacherPrompt,
5 headless::EpAppState,
6 metrics,
7 parse_output::parse_prediction_output,
8 predict::run_prediction,
9 progress::{ExampleProgress, Step},
10 reversal_tracking,
11};
12use anyhow::Context as _;
13use edit_prediction::udiff::{apply_diff_to_string, apply_diff_to_string_with_hunk_offset};
14use gpui::AsyncApp;
15use serde::Serialize;
16use std::fs::File;
17use std::io::BufWriter;
18use std::path::Path;
19use std::sync::Arc;
20
21pub async fn run_scoring(
22 example: &mut Example,
23 args: &PredictArgs,
24 app_state: Arc<EpAppState>,
25 example_progress: &ExampleProgress,
26 cx: AsyncApp,
27) -> anyhow::Result<()> {
28 run_prediction(example, args, app_state, example_progress, cx).await?;
29
30 let progress = example_progress.start(Step::Score);
31
32 progress.set_substatus("applying patches");
33 let prompt_inputs = example
34 .prompt_inputs
35 .as_ref()
36 .context("prompt_inputs is required for scoring - run prediction first or ensure JSON includes prompt_inputs")?;
37 let original_text: &str = prompt_inputs.cursor_excerpt.as_ref();
38 let expected_patches_with_cursors = example.spec.expected_patches_with_cursor_positions();
39
40 let expected_texts: Vec<String> = expected_patches_with_cursors
41 .iter()
42 .map(|(patch, _)| {
43 apply_diff_to_string(patch, original_text)
44 .with_context(|| format!("Expected patch did not apply for {}", example.spec.name))
45 })
46 .collect::<Result<Vec<_>, _>>()?;
47
48 // For Teacher prompts, we need to extract the editable region to properly compute cursor offsets.
49 // The actual_cursor_offset from Teacher is relative to the editable region, while the expected
50 // cursor from the patch is relative to the hunk. We need to apply the patch to the editable
51 // region to find where the hunk matched, then compute the expected cursor position.
52 let old_editable_region = if let Some(p) = example.prompt.as_ref() {
53 if matches!(
54 p.provider,
55 PredictionProvider::Teacher(_) | PredictionProvider::TeacherNonBatching(_)
56 ) {
57 Some(
58 TeacherPrompt::extract_editable_region(&p.input)?
59 .replace(TeacherPrompt::USER_CURSOR_MARKER, ""),
60 )
61 } else {
62 None
63 }
64 } else {
65 None
66 };
67
68 let zero_scores = ExampleScore {
69 delta_chr_f: 0.0,
70 braces_disbalance: 0,
71 exact_lines_tp: 0,
72 exact_lines_fp: 0,
73 exact_lines_fn: 0,
74 reversal_ratio: 0.0,
75 cursor_distance: None,
76 cursor_exact_match: None,
77 wrong_editable_region: None,
78 has_isolated_whitespace_changes: false,
79 inserted_tokens: 0,
80 deleted_tokens: 0,
81 };
82
83 let cursor_path = example.spec.cursor_path.as_ref();
84
85 progress.set_substatus("computing metrics");
86 let mut scores = vec![];
87 for prediction in &example.predictions {
88 let actual_patch = prediction.actual_patch.clone().or_else(|| {
89 parse_prediction_output(example, &prediction.actual_output, prediction.provider)
90 .ok()
91 .map(|(patch, _)| patch)
92 });
93
94 let Some(actual_patch) = actual_patch else {
95 scores.push(zero_scores.clone());
96 continue;
97 };
98
99 let token_changes = metrics::count_patch_token_changes(&actual_patch);
100
101 let actual_text = match apply_diff_to_string(&actual_patch, original_text) {
102 Ok(text) => text,
103 Err(_) => {
104 let mut s = zero_scores.clone();
105 s.inserted_tokens = token_changes.inserted_tokens;
106 s.deleted_tokens = token_changes.deleted_tokens;
107 scores.push(s);
108 continue;
109 }
110 };
111
112 let mut best_delta_chr_f = 0.0f32;
113 let mut best_expected_cursor: Option<usize> = None;
114 let mut best_patch_idx: Option<usize> = None;
115
116 for (idx, expected) in expected_texts.iter().enumerate() {
117 let delta_chr_f = metrics::delta_chr_f(original_text, expected, &actual_text) as f32;
118 if delta_chr_f > best_delta_chr_f {
119 best_delta_chr_f = delta_chr_f;
120 best_patch_idx = Some(idx);
121 }
122 }
123
124 if let Some(idx) = best_patch_idx {
125 // Get the raw cursor offset from the expected patch (relative to hunk new text)
126 let expected_cursor_in_patch = expected_patches_with_cursors
127 .get(idx)
128 .and_then(|(_, cursor)| *cursor);
129
130 // For Teacher prompts, we need to apply the patch to the editable region
131 // to find where the hunk matched, then compute the actual cursor position
132 if let (Some(editable_region), Some(cursor_in_patch)) =
133 (&old_editable_region, expected_cursor_in_patch)
134 {
135 let (patch, _) = &expected_patches_with_cursors[idx];
136 if let Ok((_, hunk_offset)) =
137 apply_diff_to_string_with_hunk_offset(patch, editable_region)
138 {
139 let hunk_start = hunk_offset.unwrap_or(0);
140 best_expected_cursor = Some(hunk_start + cursor_in_patch);
141 }
142 } else {
143 // For non-Teacher prompts or if we can't compute, use raw offset
144 best_expected_cursor = expected_cursor_in_patch;
145 }
146 }
147
148 let disbalance_before = metrics::braces_disbalance(&original_text);
149 let disbalance_after = metrics::braces_disbalance(&actual_text);
150 let braces_disbalance = disbalance_after.saturating_sub(disbalance_before);
151
152 // Compute exact lines match against best matching expected patch
153 let best_exact_lines = expected_patches_with_cursors
154 .iter()
155 .map(|(expected_patch, _)| metrics::exact_lines_match(expected_patch, &actual_patch))
156 .max_by_key(|m| m.true_positives)
157 .unwrap_or_default();
158
159 // Compute reversal ratio
160 let reversal_ratio = reversal_tracking::compute_prediction_reversal_ratio(
161 prompt_inputs,
162 &actual_text,
163 cursor_path,
164 );
165
166 // Compute cursor position metrics
167 let (cursor_distance, cursor_exact_match) =
168 compute_cursor_metrics(best_expected_cursor, prediction.actual_cursor.as_ref());
169
170 // Compute approximation of editable region correctness
171 let wrong_editable_region = Some(!metrics::is_editable_region_correct(&actual_patch));
172
173 // Check for isolated whitespace changes.
174 let has_isolated_whitespace_changes = metrics::has_isolated_whitespace_changes(
175 &actual_patch,
176 prediction.actual_cursor.as_ref(),
177 );
178
179 scores.push(ExampleScore {
180 delta_chr_f: best_delta_chr_f,
181 braces_disbalance,
182 exact_lines_tp: best_exact_lines.true_positives,
183 exact_lines_fp: best_exact_lines.false_positives,
184 exact_lines_fn: best_exact_lines.false_negatives,
185 reversal_ratio,
186 cursor_distance,
187 cursor_exact_match,
188 wrong_editable_region,
189 has_isolated_whitespace_changes,
190 inserted_tokens: token_changes.inserted_tokens,
191 deleted_tokens: token_changes.deleted_tokens,
192 });
193 }
194
195 example.score = scores;
196 Ok(())
197}
198
199fn compute_cursor_metrics(
200 expected_cursor_editable_region_offset: Option<usize>,
201 actual_cursor: Option<&ActualCursor>,
202) -> (Option<usize>, Option<bool>) {
203 match (expected_cursor_editable_region_offset, actual_cursor) {
204 (Some(expected), Some(actual)) => {
205 let distance = expected.abs_diff(actual.editable_region_offset.unwrap_or_default());
206 let exact_match = distance == 0;
207 (Some(distance), Some(exact_match))
208 }
209 (None, None) => {
210 // Neither has cursor position - skip cursor scoring
211 (None, None)
212 }
213 (Some(_), None) | (None, Some(_)) => {
214 // Only one has cursor position - count as miss
215 (None, Some(false))
216 }
217 }
218}
219
220pub fn print_report(examples: &[Example]) {
221 use crate::metrics::ClassificationMetrics;
222
223 const LINE_WIDTH: usize = 101;
224 let separator = "─".repeat(LINE_WIDTH);
225
226 println!("{}", separator);
227 println!(
228 "{:<40} {:>8} {:>5} {:>7} {:>7} {:>7} {:>7} {:>6} {:>5}",
229 "Example", "DeltaChrF", "Brace", "F1", "Revert", "QaRev", "QaConf", "Cursor", "WrgER"
230 );
231 println!("{}", separator);
232
233 let mut all_delta_chr_f_scores = Vec::new();
234 let mut all_reversal_ratios = Vec::new();
235 let mut braces_disbalance_sum: usize = 0;
236 let mut total_exact_lines = ClassificationMetrics::default();
237 let mut total_scores: usize = 0;
238 let mut qa_reverts_count: usize = 0;
239 let mut qa_reverts_total: usize = 0;
240 let mut qa_confidence_sum: u64 = 0;
241 let mut qa_confidence_count: usize = 0;
242 let mut cursor_exact_matches: usize = 0;
243 let mut cursor_total: usize = 0;
244 let mut cursor_distance_sum: usize = 0;
245 let mut cursor_distance_count: usize = 0;
246 let mut wrong_editable_region_count: usize = 0;
247 let mut wrong_editable_region_total: usize = 0;
248 let mut isolated_whitespace_count: usize = 0;
249 let mut patch_inserted_tokens: Vec<usize> = Vec::new();
250 let mut patch_deleted_tokens: Vec<usize> = Vec::new();
251 let mut predictions_with_patch: usize = 0;
252
253 for example in examples {
254 for (score_idx, score) in example.score.iter().enumerate() {
255 let exact_lines = ClassificationMetrics {
256 true_positives: score.exact_lines_tp,
257 false_positives: score.exact_lines_fp,
258 false_negatives: score.exact_lines_fn,
259 };
260
261 // Get QA results for this prediction if available
262 let qa_result = example.qa.get(score_idx).and_then(|q| q.as_ref());
263 let qa_reverts_str = qa_result
264 .and_then(|q| q.reverts_edits)
265 .map(|v| if v { "yes" } else { "no" })
266 .unwrap_or("-");
267 let qa_conf_str = qa_result
268 .and_then(|q| q.confidence)
269 .map(|v| format!("{}", v))
270 .unwrap_or("-".to_string());
271
272 // Format wrong editable region metric
273 let wrong_er_str = match score.wrong_editable_region {
274 Some(true) => "✗",
275 Some(false) => "",
276 None => "",
277 };
278
279 // Format cursor metric
280 let cursor_str = match (score.cursor_exact_match, score.cursor_distance) {
281 (Some(true), _) => "✓".to_string(),
282 (Some(false), Some(dist)) => format!("±{}", dist),
283 (Some(false), None) => "✗".to_string(),
284 (None, _) => "-".to_string(),
285 };
286
287 println!(
288 "{:<40} {:>8.2} {:>5} {:>6.1}% {:>6.1}% {:>7} {:>7} {:>6} {:>5}",
289 truncate_name(&example.spec.name, 40),
290 score.delta_chr_f,
291 score.braces_disbalance,
292 exact_lines.f1() * 100.0,
293 score.reversal_ratio * 100.0,
294 qa_reverts_str,
295 qa_conf_str,
296 cursor_str,
297 wrong_er_str
298 );
299
300 all_delta_chr_f_scores.push(score.delta_chr_f);
301 all_reversal_ratios.push(score.reversal_ratio);
302 total_scores += 1;
303 braces_disbalance_sum += score.braces_disbalance;
304 total_exact_lines.true_positives += score.exact_lines_tp;
305 total_exact_lines.false_positives += score.exact_lines_fp;
306 total_exact_lines.false_negatives += score.exact_lines_fn;
307
308 // Accumulate QA metrics
309 if let Some(qa) = qa_result {
310 if let Some(reverts) = qa.reverts_edits {
311 qa_reverts_total += 1;
312 if reverts {
313 qa_reverts_count += 1;
314 }
315 }
316 if let Some(conf) = qa.confidence {
317 qa_confidence_sum += conf as u64;
318 qa_confidence_count += 1;
319 }
320 }
321
322 // Accumulate wrong editable region metrics
323 if let Some(wrong) = score.wrong_editable_region {
324 wrong_editable_region_total += 1;
325 if wrong {
326 wrong_editable_region_count += 1;
327 }
328 }
329
330 // Accumulate isolated whitespace metrics
331 if score.has_isolated_whitespace_changes {
332 isolated_whitespace_count += 1;
333 }
334
335 // Accumulate token change metrics (only for predictions that produced a patch)
336 let has_patch = example
337 .predictions
338 .get(score_idx)
339 .and_then(|p| p.actual_patch.as_ref())
340 .is_some_and(|p| !p.is_empty());
341 if has_patch {
342 predictions_with_patch += 1;
343 patch_inserted_tokens.push(score.inserted_tokens);
344 patch_deleted_tokens.push(score.deleted_tokens);
345 }
346
347 // Accumulate cursor metrics
348 if let Some(exact_match) = score.cursor_exact_match {
349 cursor_total += 1;
350 if exact_match {
351 cursor_exact_matches += 1;
352 }
353 }
354 if let Some(dist) = score.cursor_distance {
355 cursor_distance_sum += dist;
356 cursor_distance_count += 1;
357 }
358 }
359 }
360
361 println!("{}", separator);
362
363 if !all_delta_chr_f_scores.is_empty() {
364 let avg_delta_chr_f: f32 =
365 all_delta_chr_f_scores.iter().sum::<f32>() / all_delta_chr_f_scores.len() as f32;
366 let avg_reversal_ratio: f32 =
367 all_reversal_ratios.iter().sum::<f32>() / all_reversal_ratios.len() as f32;
368 let braces_disbalance_avg: f32 = braces_disbalance_sum as f32 / total_scores as f32;
369
370 let qa_reverts_str = if qa_reverts_total > 0 {
371 format!(
372 "{:.1}%",
373 qa_reverts_count as f32 / qa_reverts_total as f32 * 100.0
374 )
375 } else {
376 "-".to_string()
377 };
378 let qa_conf_str = if qa_confidence_count > 0 {
379 format!(
380 "{:.1}",
381 qa_confidence_sum as f32 / qa_confidence_count as f32
382 )
383 } else {
384 "-".to_string()
385 };
386 let cursor_str = if cursor_total > 0 {
387 format!(
388 "{:.0}%",
389 cursor_exact_matches as f32 / cursor_total as f32 * 100.0
390 )
391 } else {
392 "-".to_string()
393 };
394 let wrong_er_str = if wrong_editable_region_total > 0 {
395 format!(
396 "{:.2}%",
397 wrong_editable_region_count as f32 / wrong_editable_region_total as f32 * 100.0
398 )
399 } else {
400 "-".to_string()
401 };
402 let isolated_ws_str = if total_scores > 0 {
403 format!(
404 "{}/{} ({:.1}%)",
405 isolated_whitespace_count,
406 total_scores,
407 isolated_whitespace_count as f32 / total_scores as f32 * 100.0
408 )
409 } else {
410 "-".to_string()
411 };
412 let avg_cursor_distance = if cursor_distance_count > 0 {
413 Some(cursor_distance_sum as f32 / cursor_distance_count as f32)
414 } else {
415 None
416 };
417
418 println!(
419 "{:<40} {:>8.2} {:>5.1} {:>6.1}% {:>6.1}% {:>7} {:>7} {:>6} {:>5}",
420 "TOTAL / AVERAGE",
421 avg_delta_chr_f,
422 braces_disbalance_avg,
423 total_exact_lines.f1() * 100.0,
424 avg_reversal_ratio * 100.0,
425 qa_reverts_str,
426 qa_conf_str,
427 cursor_str,
428 wrong_er_str
429 );
430 println!("{}", separator);
431
432 // Print additional cursor metrics if available
433 if let Some(avg_dist) = avg_cursor_distance {
434 println!(
435 "Cursor: {}/{} exact matches ({:.0}%), avg distance: {:.1} bytes",
436 cursor_exact_matches,
437 cursor_total,
438 cursor_exact_matches as f32 / cursor_total as f32 * 100.0,
439 avg_dist
440 );
441 }
442
443 // Print isolated whitespace metrics
444 if total_scores > 0 {
445 println!("Isolated whitespace changes: {}", isolated_ws_str);
446 }
447
448 // Print token change percentile summary (only for predictions with a patch)
449 if !patch_inserted_tokens.is_empty() {
450 patch_inserted_tokens.sort_unstable();
451 patch_deleted_tokens.sort_unstable();
452 let mut patch_total_tokens: Vec<usize> = patch_inserted_tokens
453 .iter()
454 .zip(patch_deleted_tokens.iter())
455 .map(|(i, d)| i + d)
456 .collect();
457 patch_total_tokens.sort_unstable();
458
459 let patch_rate = predictions_with_patch as f32 / total_scores as f32 * 100.0;
460 println!();
461 println!(
462 "Token changes ({}/{} predictions produced a patch, {:.1}% — table includes only those)",
463 predictions_with_patch, total_scores, patch_rate
464 );
465 println!(
466 "{:<20} {:>8} {:>8} {:>8} {:>8} {:>8}",
467 "", "p25", "p50", "p75", "p90", "p99"
468 );
469 println!("{}", "─".repeat(LINE_WIDTH));
470 println!(
471 "{:<20} {:>8} {:>8} {:>8} {:>8} {:>8}",
472 "Inserted tokens",
473 percentile(&patch_inserted_tokens, 25),
474 percentile(&patch_inserted_tokens, 50),
475 percentile(&patch_inserted_tokens, 75),
476 percentile(&patch_inserted_tokens, 90),
477 percentile(&patch_inserted_tokens, 99),
478 );
479 println!(
480 "{:<20} {:>8} {:>8} {:>8} {:>8} {:>8}",
481 "Deleted tokens",
482 percentile(&patch_deleted_tokens, 25),
483 percentile(&patch_deleted_tokens, 50),
484 percentile(&patch_deleted_tokens, 75),
485 percentile(&patch_deleted_tokens, 90),
486 percentile(&patch_deleted_tokens, 99),
487 );
488 println!(
489 "{:<20} {:>8} {:>8} {:>8} {:>8} {:>8}",
490 "Total tokens",
491 percentile(&patch_total_tokens, 25),
492 percentile(&patch_total_tokens, 50),
493 percentile(&patch_total_tokens, 75),
494 percentile(&patch_total_tokens, 90),
495 percentile(&patch_total_tokens, 99),
496 );
497 }
498 }
499
500 println!("\n");
501}
502
503fn percentile(sorted_values: &[usize], p: usize) -> usize {
504 if sorted_values.is_empty() {
505 return 0;
506 }
507 let idx = (p as f64 / 100.0 * (sorted_values.len() as f64 - 1.0)).round() as usize;
508 sorted_values[idx.min(sorted_values.len() - 1)]
509}
510
511fn truncate_name(name: &str, max_len: usize) -> String {
512 if name.len() <= max_len {
513 name.to_string()
514 } else {
515 format!("{}...", &name[..max_len - 3])
516 }
517}
518
519#[derive(Serialize)]
520pub struct SummaryJson {
521 pub total_examples: usize,
522 pub avg_delta_chr_f: f32,
523 pub avg_braces_disbalance: f32,
524 pub exact_lines_true_positives: usize,
525 pub exact_lines_false_positives: usize,
526 pub exact_lines_false_negatives: usize,
527 pub exact_lines_precision: f64,
528 pub exact_lines_recall: f64,
529 pub exact_lines_f1: f64,
530 pub avg_reversal_ratio: f32,
531 #[serde(skip_serializing_if = "Option::is_none")]
532 pub qa_avg_reverts_edits: Option<f32>,
533 #[serde(skip_serializing_if = "Option::is_none")]
534 pub qa_avg_confidence: Option<f32>,
535 #[serde(skip_serializing_if = "Option::is_none")]
536 pub cursor_exact_match_rate: Option<f32>,
537 #[serde(skip_serializing_if = "Option::is_none")]
538 pub cursor_avg_distance: Option<f32>,
539 #[serde(skip_serializing_if = "Option::is_none")]
540 pub cursor_total_evaluated: Option<usize>,
541 #[serde(skip_serializing_if = "Option::is_none")]
542 pub wrong_editable_region_rate: Option<f32>,
543 pub isolated_whitespace_rate: Option<f32>,
544}
545
546pub fn compute_summary(examples: &[Example]) -> SummaryJson {
547 use crate::metrics::ClassificationMetrics;
548
549 let mut all_delta_chr_f_scores = Vec::new();
550 let mut all_reversal_ratios = Vec::new();
551 let mut braces_disbalance_sum: usize = 0;
552 let mut total_exact_lines = ClassificationMetrics::default();
553 let mut total_scores: usize = 0;
554 let mut qa_reverts_count: usize = 0;
555 let mut qa_reverts_total: usize = 0;
556 let mut qa_confidence_sum: u64 = 0;
557 let mut qa_confidence_count: usize = 0;
558 let mut cursor_exact_matches: usize = 0;
559 let mut cursor_total: usize = 0;
560 let mut cursor_distance_sum: usize = 0;
561 let mut cursor_distance_count: usize = 0;
562 let mut wrong_editable_region_count: usize = 0;
563 let mut wrong_editable_region_total: usize = 0;
564 let mut isolated_whitespace_count: usize = 0;
565
566 for example in examples {
567 for (score_idx, score) in example.score.iter().enumerate() {
568 all_delta_chr_f_scores.push(score.delta_chr_f);
569 all_reversal_ratios.push(score.reversal_ratio);
570 total_scores += 1;
571 braces_disbalance_sum += score.braces_disbalance;
572 total_exact_lines.true_positives += score.exact_lines_tp;
573 total_exact_lines.false_positives += score.exact_lines_fp;
574 total_exact_lines.false_negatives += score.exact_lines_fn;
575
576 // Accumulate QA metrics
577 if let Some(Some(qa)) = example.qa.get(score_idx) {
578 if let Some(reverts) = qa.reverts_edits {
579 qa_reverts_total += 1;
580 if reverts {
581 qa_reverts_count += 1;
582 }
583 }
584 if let Some(conf) = qa.confidence {
585 qa_confidence_sum += conf as u64;
586 qa_confidence_count += 1;
587 }
588 }
589
590 // Accumulate wrong editable region metrics
591 if let Some(wrong) = score.wrong_editable_region {
592 wrong_editable_region_total += 1;
593 if wrong {
594 wrong_editable_region_count += 1;
595 }
596 }
597
598 // Accumulate isolated whitespace metrics
599 if score.has_isolated_whitespace_changes {
600 isolated_whitespace_count += 1;
601 }
602
603 // Accumulate cursor metrics
604 if let Some(exact_match) = score.cursor_exact_match {
605 cursor_total += 1;
606 if exact_match {
607 cursor_exact_matches += 1;
608 }
609 }
610 if let Some(dist) = score.cursor_distance {
611 cursor_distance_sum += dist;
612 cursor_distance_count += 1;
613 }
614 }
615 }
616
617 let avg_delta_chr_f = if all_delta_chr_f_scores.is_empty() {
618 0.0
619 } else {
620 all_delta_chr_f_scores.iter().sum::<f32>() / all_delta_chr_f_scores.len() as f32
621 };
622
623 let avg_reversal_ratio = if all_reversal_ratios.is_empty() {
624 0.0
625 } else {
626 all_reversal_ratios.iter().sum::<f32>() / all_reversal_ratios.len() as f32
627 };
628
629 let avg_braces_disbalance = if total_scores == 0 {
630 0.0
631 } else {
632 braces_disbalance_sum as f32 / total_scores as f32
633 };
634
635 let qa_avg_reverts_edits = if qa_reverts_total > 0 {
636 Some(qa_reverts_count as f32 / qa_reverts_total as f32)
637 } else {
638 None
639 };
640
641 let qa_avg_confidence = if qa_confidence_count > 0 {
642 Some(qa_confidence_sum as f32 / qa_confidence_count as f32)
643 } else {
644 None
645 };
646
647 let cursor_exact_match_rate = if cursor_total > 0 {
648 Some(cursor_exact_matches as f32 / cursor_total as f32)
649 } else {
650 None
651 };
652
653 let cursor_avg_distance = if cursor_distance_count > 0 {
654 Some(cursor_distance_sum as f32 / cursor_distance_count as f32)
655 } else {
656 None
657 };
658
659 let cursor_total_evaluated = if cursor_total > 0 {
660 Some(cursor_total)
661 } else {
662 None
663 };
664
665 let wrong_editable_region_rate = if wrong_editable_region_total > 0 {
666 Some(wrong_editable_region_count as f32 / wrong_editable_region_total as f32)
667 } else {
668 None
669 };
670
671 let isolated_whitespace_rate = if total_scores > 0 {
672 Some(isolated_whitespace_count as f32 / total_scores as f32)
673 } else {
674 None
675 };
676
677 SummaryJson {
678 total_examples: total_scores,
679 avg_delta_chr_f,
680 avg_braces_disbalance,
681 exact_lines_true_positives: total_exact_lines.true_positives,
682 exact_lines_false_positives: total_exact_lines.false_positives,
683 exact_lines_false_negatives: total_exact_lines.false_negatives,
684 exact_lines_precision: total_exact_lines.precision(),
685 exact_lines_recall: total_exact_lines.recall(),
686 exact_lines_f1: total_exact_lines.f1(),
687 avg_reversal_ratio,
688 qa_avg_reverts_edits,
689 qa_avg_confidence,
690 cursor_exact_match_rate,
691 cursor_avg_distance,
692 cursor_total_evaluated,
693 wrong_editable_region_rate,
694 isolated_whitespace_rate,
695 }
696}
697
698pub fn write_summary_json(examples: &[Example], path: &Path) -> anyhow::Result<()> {
699 let summary = compute_summary(examples);
700 let file = File::create(path)
701 .with_context(|| format!("Failed to create summary JSON file: {}", path.display()))?;
702 let writer = BufWriter::new(file);
703 serde_json::to_writer_pretty(writer, &summary)
704 .with_context(|| format!("Failed to write summary JSON to: {}", path.display()))?;
705 eprintln!("Wrote summary JSON to: {}", path.display());
706 Ok(())
707}