PatchJudge / patchjudge /data_loader.py
VD10's picture
Upload patchjudge/data_loader.py with huggingface_hub
0fc6b71 verified
"""Data loading pipeline for PatchJudge.
Loads SWE-bench Verified gold patches and agent-generated patches from:
1. HuggingFace datasets (AlexCuadron O1, CoderForge)
2. SWE-bench S3 bucket (139 verified agent submissions)
"""
import json
import os
import re
import logging
from pathlib import Path
from typing import Optional
from collections import defaultdict
from datasets import load_dataset
from patchjudge.models import PatchExample
logger = logging.getLogger(__name__)
class SWEBenchLoader:
"""Loads SWE-bench Verified data and agent patches."""
def __init__(self, cache_dir: str = "data"):
self.cache_dir = Path(cache_dir)
self.cache_dir.mkdir(parents=True, exist_ok=True)
self._gold_data = None # Lazy loaded
def load_gold_data(self) -> dict:
"""Load SWE-bench Verified dataset. Returns {instance_id: row_dict}."""
if self._gold_data is not None:
return self._gold_data
logger.info("Loading SWE-bench Verified dataset...")
ds = load_dataset("princeton-nlp/SWE-bench_Verified", split="test")
self._gold_data = {}
for row in ds:
self._gold_data[row["instance_id"]] = {
"instance_id": row["instance_id"],
"repo": row["repo"],
"problem_statement": row["problem_statement"],
"gold_patch": row["patch"],
"base_commit": row["base_commit"],
"test_patch": row["test_patch"],
"difficulty": row.get("difficulty", ""),
"hints_text": row.get("hints_text", ""),
}
logger.info(f"Loaded {len(self._gold_data)} SWE-bench Verified instances")
return self._gold_data
def load_coderforge_patches(self) -> list[PatchExample]:
"""Load agent patches from CoderForge (Qwen3-Coder-32B, 500 instances)."""
logger.info("Loading CoderForge agent patches...")
ds = load_dataset(
"togethercomputer/CoderForge-Preview-32B-SWE-Bench-Verified-Evaluation-trajectories",
"trajectory", split="train"
)
gold = self.load_gold_data()
examples = []
for row in ds:
# Extract instance_id from ds JSON field
try:
ds_info = json.loads(row["ds"])
instance_id = ds_info["instance_id"]
except (json.JSONDecodeError, KeyError):
# Try extracting from trajectory_id
tid = row.get("trajectory_id", "")
instance_id = tid.rsplit("_run", 1)[0] if "_run" in tid else tid
if instance_id not in gold:
continue
agent_patch = row.get("output_patch", "")
if not agent_patch or agent_patch.strip() == "":
continue
g = gold[instance_id]
ex = PatchExample(
instance_id=instance_id,
repo=g["repo"],
problem_statement=g["problem_statement"],
gold_patch=g["gold_patch"],
agent_patch=agent_patch,
agent_name="CoderForge-Qwen3-32B",
test_passed=row.get("reward", 0.0) == 1.0,
base_commit=g["base_commit"],
difficulty=g["difficulty"],
)
examples.append(ex)
logger.info(f"Loaded {len(examples)} CoderForge patches "
f"({sum(1 for e in examples if e.test_passed)} passed)")
return examples
def load_o1_patches(self) -> list[PatchExample]:
"""Load agent patches from OpenHands+O1 (500 instances)."""
logger.info("Loading OpenHands+O1 agent patches...")
ds = load_dataset(
"AlexCuadron/SWE-Bench-Verified-O1-native-tool-calling-reasoning-high-results",
split="test"
)
gold = self.load_gold_data()
examples = []
for row in ds:
issue_name = row.get("issue_name", "")
# issue_name format: "django__django-16454" — same as instance_id
instance_id = issue_name
if instance_id not in gold:
continue
agent_patch = row.get("patch", "")
if not agent_patch or agent_patch.strip() == "":
continue
g = gold[instance_id]
ex = PatchExample(
instance_id=instance_id,
repo=g["repo"],
problem_statement=g["problem_statement"],
gold_patch=g["gold_patch"],
agent_patch=agent_patch,
agent_name="OpenHands-O1-reasoning-high",
test_passed=row.get("resolved", False),
base_commit=g["base_commit"],
difficulty=g["difficulty"],
)
examples.append(ex)
logger.info(f"Loaded {len(examples)} O1 patches "
f"({sum(1 for e in examples if e.test_passed)} passed)")
return examples
def load_s3_agent_patches(
self,
agents: list[str] = None,
max_per_agent: int = 500,
) -> list[PatchExample]:
"""Load agent patches from SWE-bench S3 bucket.
Args:
agents: List of agent directory names in S3. Defaults to a curated set.
max_per_agent: Max patches per agent.
"""
try:
import boto3
from botocore import UNSIGNED
from botocore.config import Config
import requests
except ImportError:
logger.warning("boto3 not available, skipping S3 patches")
return []
if agents is None:
agents = [
"20250225_sweagent_claude-3-7-sonnet",
"20241029_OpenHands-CodeAct-2.1-sonnet-20241022",
"20241028_agentless-1.5_gpt4o",
"20241108_autocoderover-v2.0-claude-3-5-sonnet-20241022",
"20240620_sweagent_claude3.5sonnet",
]
gold = self.load_gold_data()
s3 = boto3.client('s3', config=Config(signature_version=UNSIGNED))
BUCKET = 'swe-bench-submissions'
examples = []
for agent_dir in agents:
logger.info(f"Loading patches from S3 agent: {agent_dir}")
# Get resolve labels from GitHub
resolved_ids = set()
try:
url = (
f"https://raw.githubusercontent.com/SWE-bench/experiments/"
f"main/evaluation/verified/{agent_dir}/results/results.json"
)
import requests
r = requests.get(url, timeout=10)
if r.status_code == 200:
resolved_ids = set(r.json().get("resolved", []))
logger.info(f" {agent_dir}: {len(resolved_ids)} resolved")
except Exception as e:
logger.warning(f" Could not load resolve labels for {agent_dir}: {e}")
# List instance directories
paginator = s3.get_paginator('list_objects_v2')
count = 0
try:
for page in paginator.paginate(
Bucket=BUCKET,
Prefix=f'verified/{agent_dir}/logs/',
Delimiter='/'
):
for prefix_info in page.get('CommonPrefixes', []):
if count >= max_per_agent:
break
prefix = prefix_info['Prefix']
instance_id = prefix.rstrip('/').split('/')[-1]
if instance_id not in gold:
continue
# Download patch.diff
try:
obj = s3.get_object(
Bucket=BUCKET,
Key=f'verified/{agent_dir}/logs/{instance_id}/patch.diff'
)
agent_patch = obj['Body'].read().decode('utf-8')
except Exception:
continue
if not agent_patch.strip():
continue
g = gold[instance_id]
ex = PatchExample(
instance_id=instance_id,
repo=g["repo"],
problem_statement=g["problem_statement"],
gold_patch=g["gold_patch"],
agent_patch=agent_patch,
agent_name=agent_dir,
test_passed=instance_id in resolved_ids,
base_commit=g["base_commit"],
difficulty=g["difficulty"],
)
examples.append(ex)
count += 1
except Exception as e:
logger.warning(f" Error loading from S3 for {agent_dir}: {e}")
logger.info(f" Loaded {count} patches from {agent_dir}")
logger.info(f"Total S3 patches: {len(examples)} "
f"({sum(1 for e in examples if e.test_passed)} passed)")
return examples
def build_dataset(
self,
sources: list[str] = None,
min_examples: int = 100,
include_repo_context: bool = False,
s3_agents: list[str] = None,
) -> list[PatchExample]:
"""Build the unified PatchExample dataset from multiple sources.
Args:
sources: List of sources to use. Options: 'coderforge', 'o1', 's3'.
Defaults to ['coderforge', 'o1'].
min_examples: Minimum examples to collect.
include_repo_context: If True, attempt to clone repos and gather context.
s3_agents: Agent list for S3 source.
"""
if sources is None:
sources = ["coderforge", "o1"]
all_examples = []
if "coderforge" in sources:
all_examples.extend(self.load_coderforge_patches())
if "o1" in sources:
all_examples.extend(self.load_o1_patches())
if "s3" in sources:
all_examples.extend(self.load_s3_agent_patches(agents=s3_agents))
# Deduplicate by (instance_id, agent_name)
seen = set()
unique = []
for ex in all_examples:
key = (ex.instance_id, ex.agent_name)
if key not in seen:
seen.add(key)
unique.append(ex)
logger.info(f"Total unique examples: {len(unique)} "
f"(passed: {sum(1 for e in unique if e.test_passed)}, "
f"failed: {sum(1 for e in unique if not e.test_passed)})")
if len(unique) < min_examples:
logger.warning(
f"Only {len(unique)} examples collected, "
f"below minimum of {min_examples}. "
f"Consider adding more sources."
)
return unique
def save_dataset(self, examples: list[PatchExample], filename: str = "patch_examples.jsonl"):
"""Save examples to JSONL."""
path = self.cache_dir / filename
with open(path, 'w') as f:
for ex in examples:
f.write(json.dumps(ex.to_dict()) + "\n")
logger.info(f"Saved {len(examples)} examples to {path}")
return path
def load_saved_dataset(self, filename: str = "patch_examples.jsonl") -> list[PatchExample]:
"""Load previously saved examples."""
path = self.cache_dir / filename
examples = []
with open(path) as f:
for line in f:
if line.strip():
examples.append(PatchExample.from_dict(json.loads(line)))
logger.info(f"Loaded {len(examples)} examples from {path}")
return examples
def extract_repo_context_from_diff(diff: str) -> list[str]:
"""Extract filenames mentioned in a diff."""
files = []
for line in diff.split('\n'):
if line.startswith('diff --git'):
# Extract b/path
match = re.search(r'b/(.+)$', line)
if match:
files.append(match.group(1))
elif line.startswith('---') and not line.startswith('--- /dev/null'):
match = re.search(r'a/(.+)$', line)
if match:
files.append(match.group(1))
return list(set(files))
def get_diff_stats(diff: str) -> dict:
"""Get basic stats from a unified diff."""
lines = diff.split('\n')
added = sum(1 for l in lines if l.startswith('+') and not l.startswith('+++'))
removed = sum(1 for l in lines if l.startswith('-') and not l.startswith('---'))
files = len(extract_repo_context_from_diff(diff))
hunks = sum(1 for l in lines if l.startswith('@@'))
return {
"lines_added": added,
"lines_removed": removed,
"files_changed": files,
"hunks": hunks,
}
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO)
loader = SWEBenchLoader()
# Load from HF datasets (no S3 dependency)
examples = loader.build_dataset(sources=["coderforge", "o1"])
# Stats
passed = sum(1 for e in examples if e.test_passed)
failed = len(examples) - passed
repos = set(e.repo for e in examples)
agents = set(e.agent_name for e in examples)
print(f"\n{'='*60}")
print(f"PatchJudge Dataset Summary")
print(f"{'='*60}")
print(f"Total examples: {len(examples)}")
print(f" Test passed: {passed}")
print(f" Test failed: {failed}")
print(f"Unique instances: {len(set(e.instance_id for e in examples))}")
print(f"Unique repos: {len(repos)}")
print(f"Agent sources: {agents}")
print(f"\nDifficulty distribution:")
diff_counts = defaultdict(int)
for e in examples:
diff_counts[e.difficulty] += 1
for d, c in sorted(diff_counts.items()):
print(f" {d}: {c}")
# Save
path = loader.save_dataset(examples)
print(f"\nSaved to: {path}")