File size: 3,856 Bytes
ebc85b3 | 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 | #!/usr/bin/env python3
import json
from pathlib import Path
# 扫描 qwen7b_trainset，筛选 total_pbl_loss < 0.01
base_dir = Path("/home/v-meiszhang/amlt-project/respace/eval/viz/qwen7b_trainset")
good_samples = set()
bad_samples = set()
for subdir in sorted(base_dir.iterdir()):
if not subdir.is_dir() or subdir.name.startswith('.'):
continue
metrics_file = subdir / "evaluation_metrics.json"
if not metrics_file.exists():
continue
try:
with open(metrics_file, 'r', encoding='utf-8') as f:
metrics = json.load(f)
total_pbl_loss = metrics.get('total_pbl_loss', float('inf'))
if total_pbl_loss < 0.01:
good_samples.add(subdir.name)
else:
bad_samples.add(subdir.name)
except:
pass
print(f"Good: {len(good_samples)}, Bad: {len(bad_samples)}")
# 读取训练数据
training_data_file = Path("/home/v-meiszhang/amlt-project/respace/training_data_with_content_v3.1.json")
with open(training_data_file, 'r', encoding='utf-8') as f:
raw_data = json.load(f)
# 数据可能是 {metadata, data} 结构
if isinstance(raw_data, dict) and 'data' in raw_data:
training_data = raw_data['data']
else:
training_data = raw_data
print(f"Training data: {len(training_data)}")
# 检查结构：打印第一条数据的所有字段名
if training_data:
first_item = training_data[0]
print(f"\nFields in first item: {list(first_item.keys())}")
print(f"\nFirst item sample:")
for k, v in first_item.items():
if isinstance(v, str):
print(f" {k}: {v[:80]}")
elif isinstance(v, (list, dict)):
print(f" {k}: {type(v).__name__} (len={len(v)})")
else:
print(f" {k}: {v}")
# 尝试匹配
print(f"\n{'='*60}")
print("Attempting to match samples...")
print(f"{'='*60}")
matched_count = 0
updated_count = 0
id_fields_found = {}
for i, data_item in enumerate(training_data):
candidates = [
('sample_id', data_item.get('sample_id')),
('id', data_item.get('id')),
('uuid', data_item.get('uuid')),
('scene_id', data_item.get('scene_id')),
('jid', data_item.get('jid')),
]
sample_id = None
field_name = None
for fname, fvalue in candidates:
if fvalue is not None:
sample_id = fvalue
field_name = fname
id_fields_found[fname] = id_fields_found.get(fname, 0) + 1
break
if sample_id is None:
if i < 5:
print(f"Item {i}: NO ID FOUND")
continue
if sample_id in good_samples or sample_id in bad_samples:
matched_count += 1
if sample_id in bad_samples:
data_item['split'] = 'filtered'
updated_count += 1
if i < 5:
print(f"Item {i}: MATCHED (field={field_name}, id={sample_id[:40]}...)")
print(f"\n{'='*60}")
print(f"Results:")
print(f" Matched: {matched_count} / {len(training_data)}")
print(f" Updated: {updated_count}")
print(f"\nID fields found in data:")
for fname, count in sorted(id_fields_found.items(), key=lambda x: -x[1]):
print(f" {fname}: {count}")
# 保存结果
if matched_count > 0:
output_file = training_data_file.parent / "training_data_with_content_v3.1_filtered.json"
# 保持原始结构
if isinstance(raw_data, dict) and 'data' in raw_data:
raw_data['data'] = training_data
output_data = raw_data
else:
output_data = training_data
with open(output_file, 'w', encoding='utf-8') as f:
json.dump(output_data, f, ensure_ascii=False, indent=2)
print(f"\n✓ Saved to {output_file.name}")
else:
print(f"\n✗ No matches found - check ID field names above")
print(f"\nSample from good_samples (first 5):")
for sid in sorted(good_samples)[:5]:
print(f" {sid}")
|