File size: 9,786 Bytes
3581a58 | 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 | import re
import torch
import logging
from typing import List, Dict, Optional
from nltk.translate.bleu_score import corpus_bleu, SmoothingFunction
logger = logging.getLogger(__name__)
_BERTSCORE_NUM_LAYERS_OVERRIDES = {
"models/codebert-base": 10,
"models/codebert-base-mlm": 10
}
def _resolve_bertscore_num_layers(model_type: str) -> Optional[int]:
try:
from bert_score.utils import model2layers
except Exception:
model2layers = {}
if model_type in model2layers:
return model2layers[model_type]
short_name = model_type.split("/")[-1]
if short_name in model2layers:
return model2layers[short_name]
if model_type in _BERTSCORE_NUM_LAYERS_OVERRIDES:
return _BERTSCORE_NUM_LAYERS_OVERRIDES[model_type]
name_lower = model_type.replace("\\", "/").lower()
for key in sorted(_BERTSCORE_NUM_LAYERS_OVERRIDES, key=len, reverse=True):
key_short = key.split("/")[-1].lower()
if key_short in name_lower:
return _BERTSCORE_NUM_LAYERS_OVERRIDES[key]
return None
class EvaluationMetrics:
@staticmethod
def _tokenize_sql(sql: str) -> List[str]:
"""Basic SQL tokenization for BLEU/ROUGE."""
if not sql: return []
sql = sql.lower().strip()
sql = re.sub(r'[^a-z0-9\s_\*\(\)\,\.\'\"\=\>\<\!\+\-\%]', ' ', sql)
return sql.split()
@staticmethod
def compute_bleu(references: List[str], predictions: List[str]) -> float:
try:
refs = [[EvaluationMetrics._tokenize_sql(r)] for r in references]
preds = [EvaluationMetrics._tokenize_sql(p) for p in predictions]
smoother = SmoothingFunction().method1
score = corpus_bleu(refs, preds, smoothing_function=smoother)
return round(score, 4)
except Exception as e:
logger.error(f"BLEU calculation failed: {e}")
return 0.0
@staticmethod
def compute_rouge(references: List[str], predictions: List[str]) -> Dict[str, float]:
try:
from rouge_score import rouge_scorer
scorer = rouge_scorer.RougeScorer(['rouge1', 'rougeL'], use_stemmer=True)
rouge1_scores, rougeL_scores = [], []
for ref, pred in zip(references, predictions):
scores = scorer.score(ref, pred)
rouge1_scores.append(scores['rouge1'].fmeasure)
rougeL_scores.append(scores['rougeL'].fmeasure)
return {
"ROUGE-1": round(sum(rouge1_scores) / len(rouge1_scores), 4) if rouge1_scores else 0.0,
"ROUGE-L": round(sum(rougeL_scores) / len(rougeL_scores), 4) if rougeL_scores else 0.0
}
except Exception as e:
logger.error(f"ROUGE calculation failed: {e}")
return {"ROUGE-1": 0.0, "ROUGE-L": 0.0}
@staticmethod
def compute_bertscore(references: List[str], predictions: List[str], device: str = "cpu", model_type: str = "models/codebert-base") -> float:
"""CodeBERTScore: BERTScore F1 computed with a code-aware encoder (CodeBERT by default).
NOTE: `model_type` values that aren't in bert_score's built-in
`model2layers` table (which is the case for both "microsoft/codebert-base"
and "FacebookAI/roberta-base") need an explicit `num_layers`, or the
underlying `bert_score.score()` call raises a KeyError. See
`_resolve_bertscore_num_layers` above -- this is what was previously
making CodeBERTScore silently return 0.0 for every model category.
"""
try:
from bert_score import score
valid_pairs = [(r, p) for r, p in zip(references, predictions) if r.strip() and p.strip()]
if not valid_pairs:
return 0.0
refs, preds = zip(*valid_pairs)
num_layers = _resolve_bertscore_num_layers(model_type)
score_kwargs = {"num_layers": num_layers} if num_layers is not None else {}
with torch.no_grad():
P, R, F1 = score(
list(preds), list(refs),
model_type=model_type,
verbose=False,
device=device,
**score_kwargs,
)
return round(F1.mean().item(), 4)
except Exception as e:
logger.error(f"BERTScore ({model_type}) calculation failed: {e}")
return 0.0
@staticmethod
def _find_top_level_groups(s: str) -> List[tuple]:
"""Locates every top-level `{...}`/`[...]` span in `s`, skipping over
brace/bracket characters that appear inside double-quoted string
literals (so e.g. a `$regex: "a{2,3}"` value doesn't confuse the
bracket matching)."""
groups = []
depth = 0
start = None
in_string = False
i, n = 0, len(s)
while i < n:
ch = s[i]
if in_string:
if ch == '\\':
i += 2
continue
if ch == '"':
in_string = False
i += 1
continue
if ch == '"':
in_string = True
i += 1
continue
if ch in '{[':
if depth == 0:
start = i
depth += 1
elif ch in '}]':
depth -= 1
if depth == 0 and start is not None:
groups.append((start, i + 1))
start = None
i += 1
return groups
@staticmethod
def _split_top_level(inner: str, sep: str) -> List[str]:
"""Splits `inner` on `sep`, but only at bracket-depth 0 and never
inside a double-quoted string, so nested `{...}`/`[...]`/`(...)`
and comma-containing string values aren't split incorrectly."""
parts, depth, current = [], 0, []
in_string = False
i = 0
while i < len(inner):
ch = inner[i]
if in_string:
current.append(ch)
if ch == '\\' and i + 1 < len(inner):
current.append(inner[i + 1])
i += 1
elif ch == '"':
in_string = False
i += 1
continue
if ch == '"':
in_string = True
current.append(ch)
i += 1
continue
if ch in '{[(':
depth += 1
elif ch in '}])':
depth -= 1
if ch == sep and depth == 0:
parts.append(''.join(current))
current = []
else:
current.append(ch)
i += 1
parts.append(''.join(current))
return [p for p in parts if p.strip() != '']
@classmethod
def _canon_group(cls, s: str) -> str:
s = s.strip()
if s.startswith('{') and s.endswith('}'):
inner = s[1:-1]
items = cls._split_top_level(inner, ',')
kvs = []
for item in items:
if ':' not in item:
return s # malformed for our purposes -- bail out safely
k, v = item.split(':', 1)
kvs.append((k.strip(), cls._canon_group(v.strip())))
kvs.sort(key=lambda kv: kv[0])
return '{' + ','.join(f'{k}:{v}' for k, v in kvs) + '}'
elif s.startswith('[') and s.endswith(']'):
inner = s[1:-1]
items = cls._split_top_level(inner, ',')
return '[' + ','.join(cls._canon_group(i.strip()) for i in items) + ']'
return s
@classmethod
def _canonicalize_mql(cls, s: str) -> str:
try:
if s.count('{') != s.count('}') or s.count('[') != s.count(']'):
return s
groups = cls._find_top_level_groups(s)
if not groups:
return s
out, last = [], 0
for gstart, gend in groups:
out.append(s[last:gstart])
out.append(cls._canon_group(s[gstart:gend]))
last = gend
out.append(s[last:])
return ''.join(out)
except Exception:
return s
# @classmethod
# def compute_exact_match(cls, gold: str, pred: str) -> float:
# if not gold or not pred:
# return 0.0
# def normalize(s):
# s = s.strip().rstrip(';')
# s = s.replace("'", '"') # Normalize single quotes to double quotes
# s = cls._canonicalize_mql(s) # Order-independent object-key comparison
# s = re.sub(r'\s+', '', s) # Remove all remaining whitespace
# return s.lower()
# return 1.0 if normalize(gold) == normalize(pred) else 0.0
@staticmethod
def _tokenize_mql(query: str) -> List[str]:
"""Tokenizer optimized for MongoDB Query Language (MQL) brackets and punctuation."""
if not query: return []
# Add spaces around brackets, commas, colons, and dots
query = re.sub(r'([(){}[\],:.])', r' \1 ', query)
return [tok.lower() for tok in query.split() if tok.strip()]
@staticmethod
def compute_response_accuracy(gold: str, pred: str) -> float:
"""Semantic response accuracy (recall-based token overlap)."""
if not gold or not pred: return 0.0
g_tokens = set(EvaluationMetrics._tokenize_mql(gold))
p_tokens = set(EvaluationMetrics._tokenize_mql(pred))
if not g_tokens: return 0.0
overlap = len(g_tokens.intersection(p_tokens))
return round(overlap / len(g_tokens), 4) |