1use crate::{
2 PredictArgs, PredictionProvider,
3 example::{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 original_text = &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 .content;
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 };
80
81 let prompt_inputs = example.prompt_inputs.as_ref().unwrap();
82 let cursor_path = example.spec.cursor_path.as_ref();
83
84 progress.set_substatus("computing metrics");
85 let mut scores = vec![];
86 for prediction in &example.predictions {
87 let actual_patch = prediction.actual_patch.clone().or_else(|| {
88 parse_prediction_output(example, &prediction.actual_output, prediction.provider)
89 .ok()
90 .map(|(patch, _)| patch)
91 });
92
93 let Some(actual_patch) = actual_patch else {
94 scores.push(zero_scores.clone());
95 continue;
96 };
97
98 let actual_text = match apply_diff_to_string(&actual_patch, original_text) {
99 Ok(text) => text,
100 Err(_) => {
101 scores.push(zero_scores.clone());
102 continue;
103 }
104 };
105
106 let mut best_delta_chr_f = 0.0f32;
107 let mut best_expected_cursor: Option<usize> = None;
108 let mut best_patch_idx: Option<usize> = None;
109
110 for (idx, expected) in expected_texts.iter().enumerate() {
111 let delta_chr_f = metrics::delta_chr_f(original_text, expected, &actual_text) as f32;
112 if delta_chr_f > best_delta_chr_f {
113 best_delta_chr_f = delta_chr_f;
114 best_patch_idx = Some(idx);
115 }
116 }
117
118 if let Some(idx) = best_patch_idx {
119 // Get the raw cursor offset from the expected patch (relative to hunk new text)
120 let expected_cursor_in_patch = expected_patches_with_cursors
121 .get(idx)
122 .and_then(|(_, cursor)| *cursor);
123
124 // For Teacher prompts, we need to apply the patch to the editable region
125 // to find where the hunk matched, then compute the actual cursor position
126 if let (Some(editable_region), Some(cursor_in_patch)) =
127 (&old_editable_region, expected_cursor_in_patch)
128 {
129 let (patch, _) = &expected_patches_with_cursors[idx];
130 if let Ok((_, hunk_offset)) =
131 apply_diff_to_string_with_hunk_offset(patch, editable_region)
132 {
133 let hunk_start = hunk_offset.unwrap_or(0);
134 best_expected_cursor = Some(hunk_start + cursor_in_patch);
135 }
136 } else {
137 // For non-Teacher prompts or if we can't compute, use raw offset
138 best_expected_cursor = expected_cursor_in_patch;
139 }
140 }
141
142 let disbalance_before = metrics::braces_disbalance(&original_text);
143 let disbalance_after = metrics::braces_disbalance(&actual_text);
144 let braces_disbalance = disbalance_after.saturating_sub(disbalance_before);
145
146 // Compute exact lines match against best matching expected patch
147 let best_exact_lines = expected_patches_with_cursors
148 .iter()
149 .map(|(expected_patch, _)| metrics::exact_lines_match(expected_patch, &actual_patch))
150 .max_by_key(|m| m.true_positives)
151 .unwrap_or_default();
152
153 // Compute reversal ratio
154 let reversal_ratio = reversal_tracking::compute_prediction_reversal_ratio(
155 prompt_inputs,
156 &actual_text,
157 cursor_path,
158 );
159
160 // Compute cursor position metrics
161 let (cursor_distance, cursor_exact_match) =
162 compute_cursor_metrics(best_expected_cursor, prediction.actual_cursor_offset);
163
164 // Compute approximation of editable region correctness
165 let wrong_editable_region = Some(!metrics::is_editable_region_correct(&actual_patch));
166
167 // Check for isolated whitespace changes
168 let has_isolated_whitespace_changes =
169 metrics::has_isolated_whitespace_changes(&actual_patch);
170
171 scores.push(ExampleScore {
172 delta_chr_f: best_delta_chr_f,
173 braces_disbalance,
174 exact_lines_tp: best_exact_lines.true_positives,
175 exact_lines_fp: best_exact_lines.false_positives,
176 exact_lines_fn: best_exact_lines.false_negatives,
177 reversal_ratio,
178 cursor_distance,
179 cursor_exact_match,
180 wrong_editable_region,
181 has_isolated_whitespace_changes,
182 });
183 }
184
185 example.score = scores;
186 Ok(())
187}
188
189fn compute_cursor_metrics(
190 expected_cursor: Option<usize>,
191 actual_cursor: Option<usize>,
192) -> (Option<usize>, Option<bool>) {
193 match (expected_cursor, actual_cursor) {
194 (Some(expected), Some(actual)) => {
195 let distance = expected.abs_diff(actual);
196 let exact_match = expected == actual;
197 (Some(distance), Some(exact_match))
198 }
199 (None, None) => {
200 // Neither has cursor position - skip cursor scoring
201 (None, None)
202 }
203 (Some(_), None) | (None, Some(_)) => {
204 // Only one has cursor position - count as miss
205 (None, Some(false))
206 }
207 }
208}
209
210pub fn print_report(examples: &[Example]) {
211 use crate::metrics::ClassificationMetrics;
212
213 const LINE_WIDTH: usize = 101;
214 let separator = "─".repeat(LINE_WIDTH);
215
216 println!("{}", separator);
217 println!(
218 "{:<40} {:>8} {:>5} {:>7} {:>7} {:>7} {:>7} {:>6} {:>5}",
219 "Example", "DeltaChrF", "Brace", "F1", "Revert", "QaRev", "QaConf", "Cursor", "WrgER"
220 );
221 println!("{}", separator);
222
223 let mut all_delta_chr_f_scores = Vec::new();
224 let mut all_reversal_ratios = Vec::new();
225 let mut braces_disbalance_sum: usize = 0;
226 let mut total_exact_lines = ClassificationMetrics::default();
227 let mut total_scores: usize = 0;
228 let mut qa_reverts_count: usize = 0;
229 let mut qa_reverts_total: usize = 0;
230 let mut qa_confidence_sum: u64 = 0;
231 let mut qa_confidence_count: usize = 0;
232 let mut cursor_exact_matches: usize = 0;
233 let mut cursor_total: usize = 0;
234 let mut cursor_distance_sum: usize = 0;
235 let mut cursor_distance_count: usize = 0;
236 let mut wrong_editable_region_count: usize = 0;
237 let mut wrong_editable_region_total: usize = 0;
238 let mut isolated_whitespace_count: usize = 0;
239
240 for example in examples {
241 for (score_idx, score) in example.score.iter().enumerate() {
242 let exact_lines = ClassificationMetrics {
243 true_positives: score.exact_lines_tp,
244 false_positives: score.exact_lines_fp,
245 false_negatives: score.exact_lines_fn,
246 };
247
248 // Get QA results for this prediction if available
249 let qa_result = example.qa.get(score_idx).and_then(|q| q.as_ref());
250 let qa_reverts_str = qa_result
251 .and_then(|q| q.reverts_edits)
252 .map(|v| if v { "yes" } else { "no" })
253 .unwrap_or("-");
254 let qa_conf_str = qa_result
255 .and_then(|q| q.confidence)
256 .map(|v| format!("{}", v))
257 .unwrap_or("-".to_string());
258
259 // Format wrong editable region metric
260 let wrong_er_str = match score.wrong_editable_region {
261 Some(true) => "✗",
262 Some(false) => "",
263 None => "",
264 };
265
266 // Format cursor metric
267 let cursor_str = match (score.cursor_exact_match, score.cursor_distance) {
268 (Some(true), _) => "✓".to_string(),
269 (Some(false), Some(dist)) => format!("±{}", dist),
270 (Some(false), None) => "✗".to_string(),
271 (None, _) => "-".to_string(),
272 };
273
274 println!(
275 "{:<40} {:>8.2} {:>5} {:>6.1}% {:>6.1}% {:>7} {:>7} {:>6} {:>5}",
276 truncate_name(&example.spec.name, 40),
277 score.delta_chr_f,
278 score.braces_disbalance,
279 exact_lines.f1() * 100.0,
280 score.reversal_ratio * 100.0,
281 qa_reverts_str,
282 qa_conf_str,
283 cursor_str,
284 wrong_er_str
285 );
286
287 all_delta_chr_f_scores.push(score.delta_chr_f);
288 all_reversal_ratios.push(score.reversal_ratio);
289 total_scores += 1;
290 braces_disbalance_sum += score.braces_disbalance;
291 total_exact_lines.true_positives += score.exact_lines_tp;
292 total_exact_lines.false_positives += score.exact_lines_fp;
293 total_exact_lines.false_negatives += score.exact_lines_fn;
294
295 // Accumulate QA metrics
296 if let Some(qa) = qa_result {
297 if let Some(reverts) = qa.reverts_edits {
298 qa_reverts_total += 1;
299 if reverts {
300 qa_reverts_count += 1;
301 }
302 }
303 if let Some(conf) = qa.confidence {
304 qa_confidence_sum += conf as u64;
305 qa_confidence_count += 1;
306 }
307 }
308
309 // Accumulate wrong editable region metrics
310 if let Some(wrong) = score.wrong_editable_region {
311 wrong_editable_region_total += 1;
312 if wrong {
313 wrong_editable_region_count += 1;
314 }
315 }
316
317 // Accumulate isolated whitespace metrics
318 if score.has_isolated_whitespace_changes {
319 isolated_whitespace_count += 1;
320 }
321
322 // Accumulate cursor metrics
323 if let Some(exact_match) = score.cursor_exact_match {
324 cursor_total += 1;
325 if exact_match {
326 cursor_exact_matches += 1;
327 }
328 }
329 if let Some(dist) = score.cursor_distance {
330 cursor_distance_sum += dist;
331 cursor_distance_count += 1;
332 }
333 }
334 }
335
336 println!("{}", separator);
337
338 if !all_delta_chr_f_scores.is_empty() {
339 let avg_delta_chr_f: f32 =
340 all_delta_chr_f_scores.iter().sum::<f32>() / all_delta_chr_f_scores.len() as f32;
341 let avg_reversal_ratio: f32 =
342 all_reversal_ratios.iter().sum::<f32>() / all_reversal_ratios.len() as f32;
343 let braces_disbalance_avg: f32 = braces_disbalance_sum as f32 / total_scores as f32;
344
345 let qa_reverts_str = if qa_reverts_total > 0 {
346 format!(
347 "{:.1}%",
348 qa_reverts_count as f32 / qa_reverts_total as f32 * 100.0
349 )
350 } else {
351 "-".to_string()
352 };
353 let qa_conf_str = if qa_confidence_count > 0 {
354 format!(
355 "{:.1}",
356 qa_confidence_sum as f32 / qa_confidence_count as f32
357 )
358 } else {
359 "-".to_string()
360 };
361 let cursor_str = if cursor_total > 0 {
362 format!(
363 "{:.0}%",
364 cursor_exact_matches as f32 / cursor_total as f32 * 100.0
365 )
366 } else {
367 "-".to_string()
368 };
369 let wrong_er_str = if wrong_editable_region_total > 0 {
370 format!(
371 "{:.2}%",
372 wrong_editable_region_count as f32 / wrong_editable_region_total as f32 * 100.0
373 )
374 } else {
375 "-".to_string()
376 };
377 let isolated_ws_str = if total_scores > 0 {
378 format!(
379 "{}/{} ({:.1}%)",
380 isolated_whitespace_count,
381 total_scores,
382 isolated_whitespace_count as f32 / total_scores as f32 * 100.0
383 )
384 } else {
385 "-".to_string()
386 };
387 let avg_cursor_distance = if cursor_distance_count > 0 {
388 Some(cursor_distance_sum as f32 / cursor_distance_count as f32)
389 } else {
390 None
391 };
392
393 println!(
394 "{:<40} {:>8.2} {:>5.1} {:>6.1}% {:>6.1}% {:>7} {:>7} {:>6} {:>5}",
395 "TOTAL / AVERAGE",
396 avg_delta_chr_f,
397 braces_disbalance_avg,
398 total_exact_lines.f1() * 100.0,
399 avg_reversal_ratio * 100.0,
400 qa_reverts_str,
401 qa_conf_str,
402 cursor_str,
403 wrong_er_str
404 );
405 println!("{}", separator);
406
407 // Print additional cursor metrics if available
408 if let Some(avg_dist) = avg_cursor_distance {
409 println!(
410 "Cursor: {}/{} exact matches ({:.0}%), avg distance: {:.1} bytes",
411 cursor_exact_matches,
412 cursor_total,
413 cursor_exact_matches as f32 / cursor_total as f32 * 100.0,
414 avg_dist
415 );
416 }
417
418 // Print isolated whitespace metrics
419 if total_scores > 0 {
420 println!("Isolated whitespace changes: {}", isolated_ws_str);
421 }
422 }
423
424 println!("\n");
425}
426
427fn truncate_name(name: &str, max_len: usize) -> String {
428 if name.len() <= max_len {
429 name.to_string()
430 } else {
431 format!("{}...", &name[..max_len - 3])
432 }
433}
434
435#[derive(Serialize)]
436pub struct SummaryJson {
437 pub total_examples: usize,
438 pub avg_delta_chr_f: f32,
439 pub avg_braces_disbalance: f32,
440 pub exact_lines_true_positives: usize,
441 pub exact_lines_false_positives: usize,
442 pub exact_lines_false_negatives: usize,
443 pub exact_lines_precision: f64,
444 pub exact_lines_recall: f64,
445 pub exact_lines_f1: f64,
446 pub avg_reversal_ratio: f32,
447 #[serde(skip_serializing_if = "Option::is_none")]
448 pub qa_avg_reverts_edits: Option<f32>,
449 #[serde(skip_serializing_if = "Option::is_none")]
450 pub qa_avg_confidence: Option<f32>,
451 #[serde(skip_serializing_if = "Option::is_none")]
452 pub cursor_exact_match_rate: Option<f32>,
453 #[serde(skip_serializing_if = "Option::is_none")]
454 pub cursor_avg_distance: Option<f32>,
455 #[serde(skip_serializing_if = "Option::is_none")]
456 pub cursor_total_evaluated: Option<usize>,
457 #[serde(skip_serializing_if = "Option::is_none")]
458 pub wrong_editable_region_rate: Option<f32>,
459 pub isolated_whitespace_rate: Option<f32>,
460}
461
462pub fn compute_summary(examples: &[Example]) -> SummaryJson {
463 use crate::metrics::ClassificationMetrics;
464
465 let mut all_delta_chr_f_scores = Vec::new();
466 let mut all_reversal_ratios = Vec::new();
467 let mut braces_disbalance_sum: usize = 0;
468 let mut total_exact_lines = ClassificationMetrics::default();
469 let mut total_scores: usize = 0;
470 let mut qa_reverts_count: usize = 0;
471 let mut qa_reverts_total: usize = 0;
472 let mut qa_confidence_sum: u64 = 0;
473 let mut qa_confidence_count: usize = 0;
474 let mut cursor_exact_matches: usize = 0;
475 let mut cursor_total: usize = 0;
476 let mut cursor_distance_sum: usize = 0;
477 let mut cursor_distance_count: usize = 0;
478 let mut wrong_editable_region_count: usize = 0;
479 let mut wrong_editable_region_total: usize = 0;
480 let mut isolated_whitespace_count: usize = 0;
481
482 for example in examples {
483 for (score_idx, score) in example.score.iter().enumerate() {
484 all_delta_chr_f_scores.push(score.delta_chr_f);
485 all_reversal_ratios.push(score.reversal_ratio);
486 total_scores += 1;
487 braces_disbalance_sum += score.braces_disbalance;
488 total_exact_lines.true_positives += score.exact_lines_tp;
489 total_exact_lines.false_positives += score.exact_lines_fp;
490 total_exact_lines.false_negatives += score.exact_lines_fn;
491
492 // Accumulate QA metrics
493 if let Some(Some(qa)) = example.qa.get(score_idx) {
494 if let Some(reverts) = qa.reverts_edits {
495 qa_reverts_total += 1;
496 if reverts {
497 qa_reverts_count += 1;
498 }
499 }
500 if let Some(conf) = qa.confidence {
501 qa_confidence_sum += conf as u64;
502 qa_confidence_count += 1;
503 }
504 }
505
506 // Accumulate wrong editable region metrics
507 if let Some(wrong) = score.wrong_editable_region {
508 wrong_editable_region_total += 1;
509 if wrong {
510 wrong_editable_region_count += 1;
511 }
512 }
513
514 // Accumulate isolated whitespace metrics
515 if score.has_isolated_whitespace_changes {
516 isolated_whitespace_count += 1;
517 }
518
519 // Accumulate cursor metrics
520 if let Some(exact_match) = score.cursor_exact_match {
521 cursor_total += 1;
522 if exact_match {
523 cursor_exact_matches += 1;
524 }
525 }
526 if let Some(dist) = score.cursor_distance {
527 cursor_distance_sum += dist;
528 cursor_distance_count += 1;
529 }
530 }
531 }
532
533 let avg_delta_chr_f = if all_delta_chr_f_scores.is_empty() {
534 0.0
535 } else {
536 all_delta_chr_f_scores.iter().sum::<f32>() / all_delta_chr_f_scores.len() as f32
537 };
538
539 let avg_reversal_ratio = if all_reversal_ratios.is_empty() {
540 0.0
541 } else {
542 all_reversal_ratios.iter().sum::<f32>() / all_reversal_ratios.len() as f32
543 };
544
545 let avg_braces_disbalance = if total_scores == 0 {
546 0.0
547 } else {
548 braces_disbalance_sum as f32 / total_scores as f32
549 };
550
551 let qa_avg_reverts_edits = if qa_reverts_total > 0 {
552 Some(qa_reverts_count as f32 / qa_reverts_total as f32)
553 } else {
554 None
555 };
556
557 let qa_avg_confidence = if qa_confidence_count > 0 {
558 Some(qa_confidence_sum as f32 / qa_confidence_count as f32)
559 } else {
560 None
561 };
562
563 let cursor_exact_match_rate = if cursor_total > 0 {
564 Some(cursor_exact_matches as f32 / cursor_total as f32)
565 } else {
566 None
567 };
568
569 let cursor_avg_distance = if cursor_distance_count > 0 {
570 Some(cursor_distance_sum as f32 / cursor_distance_count as f32)
571 } else {
572 None
573 };
574
575 let cursor_total_evaluated = if cursor_total > 0 {
576 Some(cursor_total)
577 } else {
578 None
579 };
580
581 let wrong_editable_region_rate = if wrong_editable_region_total > 0 {
582 Some(wrong_editable_region_count as f32 / wrong_editable_region_total as f32)
583 } else {
584 None
585 };
586
587 let isolated_whitespace_rate = if total_scores > 0 {
588 Some(isolated_whitespace_count as f32 / total_scores as f32)
589 } else {
590 None
591 };
592
593 SummaryJson {
594 total_examples: total_scores,
595 avg_delta_chr_f,
596 avg_braces_disbalance,
597 exact_lines_true_positives: total_exact_lines.true_positives,
598 exact_lines_false_positives: total_exact_lines.false_positives,
599 exact_lines_false_negatives: total_exact_lines.false_negatives,
600 exact_lines_precision: total_exact_lines.precision(),
601 exact_lines_recall: total_exact_lines.recall(),
602 exact_lines_f1: total_exact_lines.f1(),
603 avg_reversal_ratio,
604 qa_avg_reverts_edits,
605 qa_avg_confidence,
606 cursor_exact_match_rate,
607 cursor_avg_distance,
608 cursor_total_evaluated,
609 wrong_editable_region_rate,
610 isolated_whitespace_rate,
611 }
612}
613
614pub fn write_summary_json(examples: &[Example], path: &Path) -> anyhow::Result<()> {
615 let summary = compute_summary(examples);
616 let file = File::create(path)
617 .with_context(|| format!("Failed to create summary JSON file: {}", path.display()))?;
618 let writer = BufWriter::new(file);
619 serde_json::to_writer_pretty(writer, &summary)
620 .with_context(|| format!("Failed to write summary JSON to: {}", path.display()))?;
621 eprintln!("Wrote summary JSON to: {}", path.display());
622 Ok(())
623}