File size: 18,642 Bytes
dc71cad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
fine_tuning/dataset_builder.py
────────────────────────────────
Build the fine-tuning dataset from Phase 4 trajectory JSONL files.

Dataset construction strategy:
  1. Load all trajectory JSONL files from results/trajectories/
  2. Filter to high-quality instances:
       - failure_category is NOT 'unknown' (has learnable signal)
       - patch is valid (starts with --- or diff --git)
       - problem_statement is >= 20 words (enough context)
  3. Format each entry as an instruction-following pair
  4. Build hard-negative augmentation:
       - For each resolved instance, create (issue, wrong_patch) β†’ label=BAD
       - Teaches the model to distinguish correct vs. plausible-but-wrong patches
  5. Split 90/10 train/val
  6. Export as JSONL with ShareGPT / Alpaca / ChatML format options

Expected input: ~300–500 trajectory entries from a full SWE-bench Lite run
Expected output: ~800–1200 training pairs (with augmentation)

ChatML format (used by DeepSeek-Coder):
  <|im_start|>system
  You are an expert Python engineer...
  <|im_end|>
  <|im_start|>user
  ## GitHub Issue
  ...
  <|im_end|>
  <|im_start|>assistant
  --- a/path/to/file.py
  +++ b/path/to/file.py
  ...
  <|im_end|>
"""
from __future__ import annotations

import json
import logging
import random
from dataclasses import dataclass, field, asdict
from pathlib import Path
from typing import Literal, Optional

logger = logging.getLogger(__name__)

# ── Format constants ──────────────────────────────────────────────────────────

SYSTEM_PROMPT = (
    "You are an expert Python software engineer specialising in bug fixes. "
    "You will be given a GitHub issue description and the relevant source files. "
    "Your task is to generate a minimal, correct unified diff patch that fixes the issue. "
    "Output ONLY the unified diff β€” no explanations, no markdown code blocks."
)

CHATML_TEMPLATE = """\
<|im_start|>system
{system}
<|im_end|>
<|im_start|>user
{user}
<|im_end|>
<|im_start|>assistant
{assistant}
<|im_end|>"""

# ── Data types ─────────────────────────────────────────────────────────────────

@dataclass
class TrainingPair:
    system: str
    user: str
    assistant: str
    metadata: dict = field(default_factory=dict)

    def to_chatml(self) -> str:
        return CHATML_TEMPLATE.format(
            system=self.system, user=self.user, assistant=self.assistant
        )

    def to_alpaca(self) -> dict:
        return {
            "instruction": self.system + "\n\n" + self.user,
            "input": "",
            "output": self.assistant,
            "metadata": self.metadata,
        }

    def to_sharegpt(self) -> dict:
        return {
            "conversations": [
                {"from": "system", "value": self.system},
                {"from": "human",  "value": self.user},
                {"from": "gpt",    "value": self.assistant},
            ],
            "metadata": self.metadata,
        }

    def to_openai(self) -> dict:
        return {
            "messages": [
                {"role": "system",    "content": self.system},
                {"role": "user",      "content": self.user},
                {"role": "assistant", "content": self.assistant},
            ],
            "metadata": self.metadata,
        }


@dataclass
class DatasetStats:
    total_trajectories: int = 0
    after_filter: int = 0
    resolved: int = 0
    unresolved_with_category: int = 0
    augmented_pairs: int = 0
    train_size: int = 0
    val_size: int = 0
    category_counts: dict = field(default_factory=dict)
    filter_reasons: dict = field(default_factory=dict)


# ── Dataset builder ────────────────────────────────────────────────────────────

class FinetuningDatasetBuilder:
    """
    Builds a fine-tuning dataset from Phase 4 trajectory JSONL files.

    Filtering criteria (all must pass):
      - failure_category != 'unknown'
      - patch is non-empty and looks like a valid diff
      - problem_statement has >= 20 words
      - (for positive pairs) instance was eventually resolved

    Augmentation:
      - Reflection pairs: (issue + failed_attempt_context) β†’ correct_patch
        These teach the model the retry behaviour.
      - The model learns: "When tests fail with AssertionError at line X,
        the correct fix is Y" β€” generalised across many instances.
    """

    def __init__(
        self,
        trajectory_dir: Path = Path("results/trajectories"),
        output_dir: Path = Path("results/fine_tuning"),
        val_fraction: float = 0.10,
        min_problem_words: int = 20,
        max_patch_chars: int = 8000,
        seed: int = 42,
    ):
        self.trajectory_dir = Path(trajectory_dir)
        self.output_dir = Path(output_dir)
        self.val_fraction = val_fraction
        self.min_problem_words = min_problem_words
        self.max_patch_chars = max_patch_chars
        self.seed = seed
        random.seed(seed)

    def build(
        self,
        include_reflection_pairs: bool = True,
        format: Literal["chatml", "alpaca", "sharegpt", "openai"] = "chatml",
    ) -> DatasetStats:
        """
        Build and export the fine-tuning dataset.

        Args:
            include_reflection_pairs: whether to include retry/reflection pairs
            format: output format for the JSONL

        Returns:
            DatasetStats with counts and breakdown
        """
        stats = DatasetStats()

        # ── Load all trajectory files ──────────────────────────────────────
        all_entries = self._load_trajectories()
        stats.total_trajectories = len(all_entries)
        logger.info("Loaded %d trajectory entries", len(all_entries))

        # ── Filter and build pairs ─────────────────────────────────────────
        pairs: list[TrainingPair] = []
        filter_reasons: dict[str, int] = {}

        for entry in all_entries:
            reason = self._filter(entry)
            if reason:
                filter_reasons[reason] = filter_reasons.get(reason, 0) + 1
                continue

            # Build pair based on whether it was resolved
            if entry.get("resolved"):
                pair = self._build_positive_pair(entry)
                stats.resolved += 1
            else:
                # Unresolved but has known failure category
                pair = self._build_negative_pair(entry)
                if pair:
                    stats.unresolved_with_category += 1

            if pair:
                pairs.append(pair)

            cat = entry.get("failure_category", "unknown")
            stats.category_counts[cat] = stats.category_counts.get(cat, 0) + 1

        stats.after_filter = len(pairs)
        stats.filter_reasons = filter_reasons
        logger.info(
            "After filtering: %d pairs (resolved=%d, unresolved=%d)",
            len(pairs), stats.resolved, stats.unresolved_with_category
        )

        # ── Reflection pair augmentation ───────────────────────────────────
        if include_reflection_pairs:
            reflection_pairs = self._build_reflection_pairs(all_entries)
            pairs.extend(reflection_pairs)
            stats.augmented_pairs = len(reflection_pairs)
            logger.info("Added %d reflection pairs", len(reflection_pairs))

        # ── Shuffle and split ──────────────────────────────────────────────
        random.shuffle(pairs)
        n_val = max(1, int(len(pairs) * self.val_fraction))
        val_pairs = pairs[:n_val]
        train_pairs = pairs[n_val:]

        stats.train_size = len(train_pairs)
        stats.val_size = len(val_pairs)

        # ── Export ─────────────────────────────────────────────────────────
        self.output_dir.mkdir(parents=True, exist_ok=True)
        self._export(train_pairs, self.output_dir / "train.jsonl", format)
        self._export(val_pairs,   self.output_dir / "val.jsonl",   format)

        # Save stats
        stats_path = self.output_dir / "dataset_stats.json"
        stats_path.write_text(json.dumps(asdict(stats), indent=2))

        logger.info(
            "Dataset built: train=%d, val=%d β†’ %s",
            stats.train_size, stats.val_size, self.output_dir
        )
        return stats

    # ── Filtering ─────────────────────────────────────────────────────────────

    def _filter(self, entry: dict) -> Optional[str]:
        """Return a reason string if entry should be filtered, else None."""
        # Must have known failure category
        if entry.get("failure_category", "unknown") == "unknown":
            return "unknown_category"

        # Must have a non-empty patch
        patch = entry.get("patch", "").strip()
        if not patch:
            return "empty_patch"
        if not (patch.startswith("---") or patch.startswith("diff --git")):
            return "invalid_patch_format"
        if len(patch) > self.max_patch_chars:
            return "patch_too_long"

        # Must have sufficient problem statement
        problem = entry.get("problem_statement", "")
        if len(problem.strip().split()) < self.min_problem_words:
            return "problem_too_short"

        return None  # passes all filters

    # ── Pair builders ─────────────────────────────────────────────────────────

    def _build_positive_pair(self, entry: dict) -> TrainingPair:
        """Build a pair from a resolved instance."""
        user_prompt = self._build_user_prompt(
            problem_statement=entry.get("problem_statement", ""),
            localised_files=entry.get("localised_files", []),
        )
        return TrainingPair(
            system=SYSTEM_PROMPT,
            user=user_prompt,
            assistant=entry["patch"],
            metadata={
                "instance_id": entry.get("instance_id"),
                "repo": entry.get("repo"),
                "failure_category": entry.get("failure_category"),
                "pair_type": "positive",
            },
        )

    def _build_negative_pair(self, entry: dict) -> Optional[TrainingPair]:
        """
        Build a pair from an unresolved instance β€” teaches the model
        to understand WHY the patch failed and what to do instead.
        Only useful if the test output contains actionable information.
        """
        test_stdout = entry.get("test_stdout", "")
        failure_category = entry.get("failure_category", "unknown")

        # Only keep categorised failures with diagnostic info
        if failure_category == "unknown" or not test_stdout:
            return None

        # Extract actionable error context
        from agent.failure_categoriser import extract_first_error_context
        error_context = extract_first_error_context(test_stdout)

        user_prompt = self._build_user_prompt(
            problem_statement=entry.get("problem_statement", ""),
            localised_files=entry.get("localised_files", []),
            failed_patch=entry.get("patch", ""),
            failure_category=failure_category,
            error_context=error_context,
        )
        # Note: assistant still gets the original patch even though it failed
        # The model learns the (issue + error) β†’ patch_fix pattern
        return TrainingPair(
            system=SYSTEM_PROMPT,
            user=user_prompt,
            assistant=entry["patch"],
            metadata={
                "instance_id": entry.get("instance_id"),
                "pair_type": "negative_with_context",
                "failure_category": failure_category,
            },
        )

    def _build_reflection_pairs(self, all_entries: list[dict]) -> list[TrainingPair]:
        """
        Build reflection pairs: (issue + attempt_k_failure) β†’ attempt_{k+1}_patch.

        For multi-attempt instances where the agent eventually succeeds,
        we pair each failed attempt with the final successful patch.
        This directly teaches the reflection behaviour.
        """
        pairs = []
        # Group by instance_id
        by_instance: dict[str, list[dict]] = {}
        for e in all_entries:
            iid = e.get("instance_id", "")
            by_instance.setdefault(iid, []).append(e)

        for iid, entries in by_instance.items():
            entries_sorted = sorted(entries, key=lambda x: x.get("attempt", 1))
            # Find final successful patch
            final = next((e for e in reversed(entries_sorted) if e.get("resolved")), None)
            if not final or not final.get("patch"):
                continue

            # Each failed attempt before the success becomes a reflection pair
            for failed_entry in entries_sorted[:-1]:
                if failed_entry.get("resolved"):
                    continue
                if self._filter(failed_entry):
                    continue

                from agent.failure_categoriser import extract_first_error_context
                error_ctx = extract_first_error_context(failed_entry.get("test_stdout", ""))

                user_prompt = self._build_user_prompt(
                    problem_statement=failed_entry.get("problem_statement", ""),
                    localised_files=failed_entry.get("localised_files", []),
                    failed_patch=failed_entry.get("patch", ""),
                    failure_category=failed_entry.get("failure_category", ""),
                    error_context=error_ctx,
                )
                pairs.append(TrainingPair(
                    system=SYSTEM_PROMPT,
                    user=user_prompt,
                    assistant=final["patch"],   # correct final patch
                    metadata={
                        "instance_id": iid,
                        "pair_type": "reflection",
                        "attempt": failed_entry.get("attempt"),
                    },
                ))

        logger.info("Generated %d reflection pairs", len(pairs))
        return pairs

    # ── Helpers ───────────────────────────────────────────────────────────────

    def _build_user_prompt(
        self,
        problem_statement: str,
        localised_files: list[str],
        failed_patch: str = "",
        failure_category: str = "",
        error_context: str = "",
    ) -> str:
        parts = [f"## GitHub Issue\n{problem_statement[:1000]}"]

        if localised_files:
            file_list = "\n".join(f"- {fp}" for fp in localised_files[:8])
            parts.append(f"## Relevant Files\n{file_list}")

        if failed_patch and failure_category:
            parts.append(
                f"## Previous Attempt Failed\n"
                f"Failure category: **{failure_category}**\n\n"
                f"```\n{error_context[:500]}\n```\n\n"
                f"Previous patch:\n```diff\n{failed_patch[:800]}\n```"
            )

        parts.append("Generate a unified diff patch that fixes the issue.")
        return "\n\n".join(parts)

    def _load_trajectories(self) -> list[dict]:
        """Load all trajectory entries from JSONL files in trajectory_dir."""
        from agent.trajectory_logger import TrajectoryLogger
        import dataclasses

        all_entries: list[dict] = []
        if not self.trajectory_dir.exists():
            logger.warning("Trajectory directory not found: %s", self.trajectory_dir)
            return all_entries

        for jsonl_path in self.trajectory_dir.glob("*.jsonl"):
            tl = TrajectoryLogger(jsonl_path)
            for entry in tl.load_all():
                all_entries.append(dataclasses.asdict(entry))

        logger.info("Loaded %d entries from %d files", len(all_entries),
                    len(list(self.trajectory_dir.glob("*.jsonl"))))
        return all_entries

    def _export(
        self,
        pairs: list[TrainingPair],
        path: Path,
        format: str,
    ) -> None:
        path.parent.mkdir(parents=True, exist_ok=True)
        with path.open("w") as f:
            for pair in pairs:
                if format == "chatml":
                    f.write(json.dumps({"text": pair.to_chatml(), "metadata": pair.metadata}) + "\n")
                elif format == "alpaca":
                    f.write(json.dumps(pair.to_alpaca()) + "\n")
                elif format == "sharegpt":
                    f.write(json.dumps(pair.to_sharegpt()) + "\n")
                elif format == "openai":
                    f.write(json.dumps(pair.to_openai()) + "\n")
        logger.info("Exported %d %s pairs to %s", len(pairs), format, path)


# ── Token count estimator ─────────────────────────────────────────────────────

def estimate_token_counts(dataset_path: Path) -> dict:
    """
    Estimate token counts for training cost estimation.
    Uses simple word-count heuristic (1 word β‰ˆ 1.3 tokens).
    """
    if not dataset_path.exists():
        return {}

    total_chars = 0
    n_pairs = 0
    with dataset_path.open() as f:
        for line in f:
            obj = json.loads(line)
            text = obj.get("text") or str(obj)
            total_chars += len(text)
            n_pairs += 1

    estimated_tokens = int(total_chars / 4)  # ~4 chars per token
    return {
        "n_pairs": n_pairs,
        "estimated_tokens": estimated_tokens,
        "estimated_tokens_per_pair": estimated_tokens // max(n_pairs, 1),
        "estimated_training_cost_usd": estimated_tokens / 1e6 * 0.12,  # rough A100 estimate
    }