Spaces:
Sleeping
Sleeping
Rename src/evaluation.py to src/validation.py
Browse files- src/evaluation.py +0 -413
- src/validation.py +274 -0
src/evaluation.py
DELETED
@@ -1,413 +0,0 @@
|
|
1 |
-
# src/evaluation.py
|
2 |
-
import torch
|
3 |
-
import numpy as np
|
4 |
-
from tqdm.auto import tqdm
|
5 |
-
from sacrebleu.metrics import BLEU, CHRF
|
6 |
-
from rouge_score import rouge_scorer
|
7 |
-
import Levenshtein
|
8 |
-
from collections import defaultdict
|
9 |
-
from transformers.models.whisper.english_normalizer import BasicTextNormalizer
|
10 |
-
import salt.constants
|
11 |
-
import datetime
|
12 |
-
import os
|
13 |
-
from google.cloud import translate_v3
|
14 |
-
from config import GOOGLE_LANG_MAP
|
15 |
-
|
16 |
-
def setup_google_translate():
|
17 |
-
"""Setup Google Cloud Translation client if credentials available."""
|
18 |
-
try:
|
19 |
-
# Check if running in HF Space with credentials
|
20 |
-
if os.getenv("GOOGLE_APPLICATION_CREDENTIALS") or os.getenv("GOOGLE_CLOUD_PROJECT"):
|
21 |
-
client = translate_v3.TranslationServiceClient()
|
22 |
-
project_id = os.getenv("GOOGLE_CLOUD_PROJECT", "sb-gcp-project-01")
|
23 |
-
parent = f"projects/{project_id}/locations/global"
|
24 |
-
return client, parent
|
25 |
-
else:
|
26 |
-
print("Google Cloud credentials not found. Google Translate will not be available.")
|
27 |
-
return None, None
|
28 |
-
except Exception as e:
|
29 |
-
print(f"Error setting up Google Translate: {e}")
|
30 |
-
return None, None
|
31 |
-
|
32 |
-
def google_translate_batch(texts, source_langs, target_langs, client, parent):
|
33 |
-
"""Translate using Google Cloud Translation API."""
|
34 |
-
translations = []
|
35 |
-
|
36 |
-
for text, src_lang, tgt_lang in tqdm(zip(texts, source_langs, target_langs),
|
37 |
-
total=len(texts), desc="Google Translate"):
|
38 |
-
try:
|
39 |
-
# Map SALT language codes to Google's format
|
40 |
-
src_google = GOOGLE_LANG_MAP.get(src_lang, src_lang)
|
41 |
-
tgt_google = GOOGLE_LANG_MAP.get(tgt_lang, tgt_lang)
|
42 |
-
|
43 |
-
# Check if language pair is supported
|
44 |
-
supported_langs = ['lg', 'ach', 'sw', 'en']
|
45 |
-
if src_google not in supported_langs or tgt_google not in supported_langs:
|
46 |
-
translations.append(f"[UNSUPPORTED: {src_lang}->{tgt_lang}]")
|
47 |
-
continue
|
48 |
-
|
49 |
-
# Make translation request
|
50 |
-
request = {
|
51 |
-
"parent": parent,
|
52 |
-
"contents": [text],
|
53 |
-
"mime_type": "text/plain",
|
54 |
-
"source_language_code": src_google,
|
55 |
-
"target_language_code": tgt_google,
|
56 |
-
}
|
57 |
-
|
58 |
-
response = client.translate_text(request=request)
|
59 |
-
translation = response.translations[0].translated_text
|
60 |
-
translations.append(translation)
|
61 |
-
|
62 |
-
except Exception as e:
|
63 |
-
print(f"Error translating '{text}': {e}")
|
64 |
-
translations.append(f"[ERROR: {str(e)[:50]}]")
|
65 |
-
|
66 |
-
return translations
|
67 |
-
|
68 |
-
def get_translation_function(model, tokenizer, model_path):
|
69 |
-
"""Get appropriate translation function based on model type."""
|
70 |
-
|
71 |
-
if model_path == 'google-translate':
|
72 |
-
client, parent = setup_google_translate()
|
73 |
-
if client is None:
|
74 |
-
raise Exception("Google Translate credentials not available")
|
75 |
-
|
76 |
-
def translation_fn(texts, from_langs, to_langs):
|
77 |
-
return google_translate_batch(texts, from_langs, to_langs, client, parent)
|
78 |
-
|
79 |
-
return translation_fn
|
80 |
-
|
81 |
-
elif 'gemma' in str(type(model)).lower() or 'gemma' in model_path.lower():
|
82 |
-
return get_gemma_translation_fn(model, tokenizer)
|
83 |
-
|
84 |
-
elif hasattr(model, 'base_model') and hasattr(model.base_model, 'model') and 'Qwen2ForCausalLM' in str(type(model.base_model.model)):
|
85 |
-
return get_qwen_translation_fn(model, tokenizer)
|
86 |
-
|
87 |
-
elif 'm2m_100' in str(type(model)).lower():
|
88 |
-
return get_nllb_translation_fn(model, tokenizer)
|
89 |
-
|
90 |
-
elif hasattr(model, 'base_model') and hasattr(model.base_model, 'model') and 'LlamaForCausalLM' in str(type(model.base_model.model)):
|
91 |
-
return get_llama_translation_fn(model, tokenizer)
|
92 |
-
|
93 |
-
else:
|
94 |
-
# Generic function for other models
|
95 |
-
return get_generic_translation_fn(model, tokenizer)
|
96 |
-
|
97 |
-
def get_gemma_translation_fn(model, tokenizer):
|
98 |
-
"""Translation function for Gemma models."""
|
99 |
-
def translation_fn(texts, from_langs, to_langs):
|
100 |
-
SYSTEM_MESSAGE = 'You are a linguist and translation assistant specialising in Ugandan languages.'
|
101 |
-
translations = []
|
102 |
-
batch_size = 4
|
103 |
-
device = next(model.parameters()).device
|
104 |
-
|
105 |
-
instructions = [
|
106 |
-
f'Translate from {salt.constants.SALT_LANGUAGE_NAMES[from_lang]} '
|
107 |
-
f'to {salt.constants.SALT_LANGUAGE_NAMES[to_lang]}: {text}'
|
108 |
-
for text, from_lang, to_lang in zip(texts, from_langs, to_langs)
|
109 |
-
]
|
110 |
-
|
111 |
-
for i in tqdm(range(0, len(instructions), batch_size), desc="Generating translations"):
|
112 |
-
batch_instructions = instructions[i:i + batch_size]
|
113 |
-
messages_list = [
|
114 |
-
[
|
115 |
-
{"role": "system", "content": SYSTEM_MESSAGE},
|
116 |
-
{"role": "user", "content": instruction}
|
117 |
-
] for instruction in batch_instructions
|
118 |
-
]
|
119 |
-
|
120 |
-
prompts = [
|
121 |
-
tokenizer.apply_chat_template(
|
122 |
-
messages, tokenize=False, add_generation_prompt=True
|
123 |
-
) for messages in messages_list
|
124 |
-
]
|
125 |
-
|
126 |
-
inputs = tokenizer(
|
127 |
-
prompts, return_tensors="pt",
|
128 |
-
padding=True, padding_side='left',
|
129 |
-
max_length=512, truncation=True
|
130 |
-
).to(device)
|
131 |
-
|
132 |
-
with torch.no_grad():
|
133 |
-
outputs = model.generate(
|
134 |
-
**inputs,
|
135 |
-
max_new_tokens=100,
|
136 |
-
temperature=0.5,
|
137 |
-
num_beams=5,
|
138 |
-
do_sample=True,
|
139 |
-
no_repeat_ngram_size=5,
|
140 |
-
pad_token_id=tokenizer.eos_token_id
|
141 |
-
)
|
142 |
-
|
143 |
-
for j in range(len(outputs)):
|
144 |
-
translation = tokenizer.decode(
|
145 |
-
outputs[j, inputs['input_ids'].shape[1]:],
|
146 |
-
skip_special_tokens=True
|
147 |
-
)
|
148 |
-
translations.append(translation)
|
149 |
-
|
150 |
-
return translations
|
151 |
-
|
152 |
-
return translation_fn
|
153 |
-
|
154 |
-
def get_qwen_translation_fn(model, tokenizer):
|
155 |
-
"""Translation function for Qwen models."""
|
156 |
-
def translation_fn(texts, from_langs, to_langs):
|
157 |
-
SYSTEM_MESSAGE = 'You are a Ugandan language assistant.'
|
158 |
-
translations = []
|
159 |
-
batch_size = 8
|
160 |
-
device = next(model.parameters()).device
|
161 |
-
|
162 |
-
instructions = [
|
163 |
-
f'Translate from {salt.constants.SALT_LANGUAGE_NAMES.get(from_lang, from_lang)} '
|
164 |
-
f'to {salt.constants.SALT_LANGUAGE_NAMES.get(to_lang, to_lang)}: {text}'
|
165 |
-
for text, from_lang, to_lang in zip(texts, from_langs, to_langs)
|
166 |
-
]
|
167 |
-
|
168 |
-
for i in tqdm(range(0, len(instructions), batch_size), desc="Generating translations"):
|
169 |
-
batch_instructions = instructions[i:i + batch_size]
|
170 |
-
messages_list = [
|
171 |
-
[
|
172 |
-
{"role": "system", "content": SYSTEM_MESSAGE},
|
173 |
-
{"role": "user", "content": instruction}
|
174 |
-
] for instruction in batch_instructions
|
175 |
-
]
|
176 |
-
|
177 |
-
prompts = [
|
178 |
-
tokenizer.apply_chat_template(
|
179 |
-
messages, tokenize=False, add_generation_prompt=True
|
180 |
-
) for messages in messages_list
|
181 |
-
]
|
182 |
-
|
183 |
-
inputs = tokenizer(
|
184 |
-
prompts, return_tensors="pt",
|
185 |
-
padding=True, padding_side='left', truncation=True
|
186 |
-
).to(device)
|
187 |
-
|
188 |
-
with torch.no_grad():
|
189 |
-
outputs = model.generate(
|
190 |
-
**inputs, max_new_tokens=100,
|
191 |
-
temperature=0.01,
|
192 |
-
pad_token_id=tokenizer.eos_token_id
|
193 |
-
)
|
194 |
-
|
195 |
-
for j in range(len(outputs)):
|
196 |
-
translation = tokenizer.decode(
|
197 |
-
outputs[j, inputs['input_ids'].shape[1]:],
|
198 |
-
skip_special_tokens=True
|
199 |
-
)
|
200 |
-
translations.append(translation)
|
201 |
-
|
202 |
-
return translations
|
203 |
-
|
204 |
-
return translation_fn
|
205 |
-
|
206 |
-
def get_nllb_translation_fn(model, tokenizer):
|
207 |
-
"""Translation function for NLLB models."""
|
208 |
-
def translation_fn(texts, source_langs, target_langs):
|
209 |
-
translations = []
|
210 |
-
language_tokens = salt.constants.SALT_LANGUAGE_TOKENS_NLLB_TRANSLATION
|
211 |
-
device = next(model.parameters()).device
|
212 |
-
|
213 |
-
for text, source_language, target_language in tqdm(
|
214 |
-
zip(texts, source_langs, target_langs), total=len(texts), desc="NLLB Translation"):
|
215 |
-
|
216 |
-
inputs = tokenizer(text, return_tensors="pt").to(device)
|
217 |
-
inputs['input_ids'][0][0] = language_tokens[source_language]
|
218 |
-
|
219 |
-
with torch.no_grad():
|
220 |
-
translated_tokens = model.generate(
|
221 |
-
**inputs,
|
222 |
-
forced_bos_token_id=language_tokens[target_language],
|
223 |
-
max_length=100,
|
224 |
-
num_beams=5,
|
225 |
-
)
|
226 |
-
|
227 |
-
result = tokenizer.batch_decode(
|
228 |
-
translated_tokens, skip_special_tokens=True)[0]
|
229 |
-
translations.append(result)
|
230 |
-
|
231 |
-
return translations
|
232 |
-
|
233 |
-
return translation_fn
|
234 |
-
|
235 |
-
def get_llama_translation_fn(model, tokenizer):
|
236 |
-
"""Translation function for Llama models."""
|
237 |
-
def translation_fn(texts, from_langs, to_langs):
|
238 |
-
DATE_TODAY = datetime.datetime.now().strftime("%d %b %Y")
|
239 |
-
SYSTEM_MESSAGE = ''
|
240 |
-
translations = []
|
241 |
-
batch_size = 8
|
242 |
-
device = next(model.parameters()).device
|
243 |
-
|
244 |
-
instructions = [
|
245 |
-
f'Translate from {salt.constants.SALT_LANGUAGE_NAMES.get(from_lang, from_lang)} '
|
246 |
-
f'to {salt.constants.SALT_LANGUAGE_NAMES.get(to_lang, to_lang)}: {text}'
|
247 |
-
for text, from_lang, to_lang in zip(texts, from_langs, to_langs)
|
248 |
-
]
|
249 |
-
|
250 |
-
for i in tqdm(range(0, len(instructions), batch_size), desc="Llama Translation"):
|
251 |
-
batch_instructions = instructions[i:i + batch_size]
|
252 |
-
messages_list = [
|
253 |
-
[
|
254 |
-
{"role": "system", "content": SYSTEM_MESSAGE},
|
255 |
-
{"role": "user", "content": instruction}
|
256 |
-
] for instruction in batch_instructions
|
257 |
-
]
|
258 |
-
|
259 |
-
prompts = [
|
260 |
-
tokenizer.apply_chat_template(
|
261 |
-
messages, tokenize=False, add_generation_prompt=True,
|
262 |
-
date_string=DATE_TODAY,
|
263 |
-
) for messages in messages_list
|
264 |
-
]
|
265 |
-
|
266 |
-
inputs = tokenizer(
|
267 |
-
prompts, return_tensors="pt",
|
268 |
-
padding=True, padding_side='left',
|
269 |
-
).to(device)
|
270 |
-
|
271 |
-
with torch.no_grad():
|
272 |
-
outputs = model.generate(
|
273 |
-
**inputs, max_new_tokens=100,
|
274 |
-
temperature=0.01,
|
275 |
-
pad_token_id=tokenizer.eos_token_id
|
276 |
-
)
|
277 |
-
|
278 |
-
for j in range(len(outputs)):
|
279 |
-
translation = tokenizer.decode(
|
280 |
-
outputs[j, inputs['input_ids'].shape[1]:],
|
281 |
-
skip_special_tokens=True
|
282 |
-
)
|
283 |
-
translations.append(translation)
|
284 |
-
|
285 |
-
return translations
|
286 |
-
|
287 |
-
return translation_fn
|
288 |
-
|
289 |
-
def get_generic_translation_fn(model, tokenizer):
|
290 |
-
"""Generic translation function for unknown model types."""
|
291 |
-
def translation_fn(texts, from_langs, to_langs):
|
292 |
-
translations = []
|
293 |
-
device = next(model.parameters()).device
|
294 |
-
|
295 |
-
for text, from_lang, to_lang in tqdm(zip(texts, from_langs, to_langs),
|
296 |
-
desc="Generic Translation"):
|
297 |
-
prompt = f"Translate from {from_lang} to {to_lang}: {text}"
|
298 |
-
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
299 |
-
|
300 |
-
with torch.no_grad():
|
301 |
-
outputs = model.generate(
|
302 |
-
**inputs,
|
303 |
-
max_new_tokens=100,
|
304 |
-
temperature=0.7,
|
305 |
-
pad_token_id=tokenizer.eos_token_id
|
306 |
-
)
|
307 |
-
|
308 |
-
translation = tokenizer.decode(
|
309 |
-
outputs[0, inputs['input_ids'].shape[1]:],
|
310 |
-
skip_special_tokens=True
|
311 |
-
)
|
312 |
-
translations.append(translation)
|
313 |
-
|
314 |
-
return translations
|
315 |
-
|
316 |
-
return translation_fn
|
317 |
-
|
318 |
-
def calculate_metrics(reference: str, prediction: str) -> dict:
|
319 |
-
"""Calculate multiple translation quality metrics."""
|
320 |
-
bleu = BLEU(effective_order=True)
|
321 |
-
bleu_score = bleu.sentence_score(prediction, [reference]).score
|
322 |
-
|
323 |
-
chrf = CHRF()
|
324 |
-
chrf_score = chrf.sentence_score(prediction, [reference]).score / 100.0
|
325 |
-
|
326 |
-
cer = Levenshtein.distance(reference, prediction) / max(len(reference), 1)
|
327 |
-
|
328 |
-
ref_words = reference.split()
|
329 |
-
pred_words = prediction.split()
|
330 |
-
wer = Levenshtein.distance(ref_words, pred_words) / max(len(ref_words), 1)
|
331 |
-
|
332 |
-
len_ratio = len(prediction) / max(len(reference), 1)
|
333 |
-
|
334 |
-
metrics = {
|
335 |
-
"bleu": bleu_score,
|
336 |
-
"chrf": chrf_score,
|
337 |
-
"cer": cer,
|
338 |
-
"wer": wer,
|
339 |
-
"len_ratio": len_ratio,
|
340 |
-
}
|
341 |
-
|
342 |
-
try:
|
343 |
-
scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)
|
344 |
-
rouge_scores = scorer.score(reference, prediction)
|
345 |
-
|
346 |
-
metrics["rouge1"] = rouge_scores['rouge1'].fmeasure
|
347 |
-
metrics["rouge2"] = rouge_scores['rouge2'].fmeasure
|
348 |
-
metrics["rougeL"] = rouge_scores['rougeL'].fmeasure
|
349 |
-
|
350 |
-
metrics["quality_score"] = (
|
351 |
-
bleu_score/100 +
|
352 |
-
chrf_score +
|
353 |
-
(1-cer) +
|
354 |
-
(1-wer) +
|
355 |
-
rouge_scores['rouge1'].fmeasure +
|
356 |
-
rouge_scores['rougeL'].fmeasure
|
357 |
-
) / 6
|
358 |
-
except Exception as e:
|
359 |
-
print(f"Error calculating ROUGE metrics: {e}")
|
360 |
-
metrics["quality_score"] = (bleu_score/100 + chrf_score + (1-cer) + (1-wer)) / 4
|
361 |
-
|
362 |
-
return metrics
|
363 |
-
|
364 |
-
def evaluate_model_full(model, tokenizer, model_path: str, test_data) -> dict:
|
365 |
-
"""Complete model evaluation pipeline."""
|
366 |
-
|
367 |
-
# Get translation function
|
368 |
-
translation_fn = get_translation_function(model, tokenizer, model_path)
|
369 |
-
|
370 |
-
# Generate predictions
|
371 |
-
print("Generating translations...")
|
372 |
-
predictions = translation_fn(
|
373 |
-
list(test_data['source']),
|
374 |
-
list(test_data['source.language']),
|
375 |
-
list(test_data['target.language']),
|
376 |
-
)
|
377 |
-
|
378 |
-
# Calculate metrics by language pair
|
379 |
-
print("Calculating metrics...")
|
380 |
-
translation_subsets = defaultdict(list)
|
381 |
-
for idx, row in test_data.iterrows():
|
382 |
-
direction = row['source.language'] + '_to_' + row['target.language']
|
383 |
-
row_dict = dict(row)
|
384 |
-
row_dict['prediction'] = predictions[idx]
|
385 |
-
translation_subsets[direction].append(row_dict)
|
386 |
-
|
387 |
-
normalizer = BasicTextNormalizer()
|
388 |
-
grouped_metrics = defaultdict(dict)
|
389 |
-
|
390 |
-
for subset in translation_subsets.keys():
|
391 |
-
subset_metrics = defaultdict(list)
|
392 |
-
for example in translation_subsets[subset]:
|
393 |
-
prediction = normalizer(str(example['prediction']))
|
394 |
-
reference = normalizer(example['target'])
|
395 |
-
metrics = calculate_metrics(reference, prediction)
|
396 |
-
for m in metrics.keys():
|
397 |
-
subset_metrics[m].append(metrics[m])
|
398 |
-
|
399 |
-
for m in subset_metrics.keys():
|
400 |
-
if subset_metrics[m]: # Check if list is not empty
|
401 |
-
grouped_metrics[subset][m] = float(np.mean(subset_metrics[m]))
|
402 |
-
|
403 |
-
# Calculate overall averages
|
404 |
-
all_metrics = list(grouped_metrics.values())[0].keys() if grouped_metrics else []
|
405 |
-
for m in all_metrics:
|
406 |
-
metric_values = []
|
407 |
-
for subset in translation_subsets.keys():
|
408 |
-
if m in grouped_metrics[subset]:
|
409 |
-
metric_values.append(grouped_metrics[subset][m])
|
410 |
-
if metric_values:
|
411 |
-
grouped_metrics['averages'][m] = float(np.mean(metric_values))
|
412 |
-
|
413 |
-
return dict(grouped_metrics)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/validation.py
ADDED
@@ -0,0 +1,274 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# src/validation.py
|
2 |
+
import pandas as pd
|
3 |
+
import numpy as np
|
4 |
+
from typing import Dict, List, Tuple, Optional
|
5 |
+
import json
|
6 |
+
import io
|
7 |
+
from config import PREDICTION_FORMAT
|
8 |
+
|
9 |
+
def validate_file_format(file_content: bytes, filename: str) -> Dict:
|
10 |
+
"""Validate uploaded file format and structure."""
|
11 |
+
|
12 |
+
try:
|
13 |
+
# Determine file type
|
14 |
+
if filename.endswith('.csv'):
|
15 |
+
df = pd.read_csv(io.BytesIO(file_content))
|
16 |
+
elif filename.endswith('.tsv'):
|
17 |
+
df = pd.read_csv(io.BytesIO(file_content), sep='\t')
|
18 |
+
elif filename.endswith('.json'):
|
19 |
+
data = json.loads(file_content.decode('utf-8'))
|
20 |
+
df = pd.DataFrame(data)
|
21 |
+
else:
|
22 |
+
return {
|
23 |
+
'valid': False,
|
24 |
+
'error': f"Unsupported file type. Use: {', '.join(PREDICTION_FORMAT['file_types'])}"
|
25 |
+
}
|
26 |
+
|
27 |
+
# Check required columns
|
28 |
+
missing_cols = set(PREDICTION_FORMAT['required_columns']) - set(df.columns)
|
29 |
+
if missing_cols:
|
30 |
+
return {
|
31 |
+
'valid': False,
|
32 |
+
'error': f"Missing required columns: {', '.join(missing_cols)}"
|
33 |
+
}
|
34 |
+
|
35 |
+
# Basic data validation
|
36 |
+
if len(df) == 0:
|
37 |
+
return {
|
38 |
+
'valid': False,
|
39 |
+
'error': "File is empty"
|
40 |
+
}
|
41 |
+
|
42 |
+
# Check for required data
|
43 |
+
if df['sample_id'].isna().any():
|
44 |
+
return {
|
45 |
+
'valid': False,
|
46 |
+
'error': "Missing sample_id values found"
|
47 |
+
}
|
48 |
+
|
49 |
+
if df['prediction'].isna().any():
|
50 |
+
na_count = df['prediction'].isna().sum()
|
51 |
+
return {
|
52 |
+
'valid': False,
|
53 |
+
'error': f"Missing prediction values found ({na_count} empty predictions)"
|
54 |
+
}
|
55 |
+
|
56 |
+
# Check for duplicates
|
57 |
+
duplicates = df['sample_id'].duplicated()
|
58 |
+
if duplicates.any():
|
59 |
+
dup_count = duplicates.sum()
|
60 |
+
return {
|
61 |
+
'valid': False,
|
62 |
+
'error': f"Duplicate sample_id values found ({dup_count} duplicates)"
|
63 |
+
}
|
64 |
+
|
65 |
+
return {
|
66 |
+
'valid': True,
|
67 |
+
'dataframe': df,
|
68 |
+
'row_count': len(df),
|
69 |
+
'columns': list(df.columns)
|
70 |
+
}
|
71 |
+
|
72 |
+
except Exception as e:
|
73 |
+
return {
|
74 |
+
'valid': False,
|
75 |
+
'error': f"Error parsing file: {str(e)}"
|
76 |
+
}
|
77 |
+
|
78 |
+
def validate_predictions_content(predictions: pd.DataFrame) -> Dict:
|
79 |
+
"""Validate prediction content quality."""
|
80 |
+
|
81 |
+
issues = []
|
82 |
+
warnings = []
|
83 |
+
|
84 |
+
# Check prediction text quality
|
85 |
+
empty_predictions = predictions['prediction'].str.strip().eq('').sum()
|
86 |
+
if empty_predictions > 0:
|
87 |
+
issues.append(f"{empty_predictions} empty predictions found")
|
88 |
+
|
89 |
+
# Check for suspiciously short predictions
|
90 |
+
short_predictions = (predictions['prediction'].str.len() < 3).sum()
|
91 |
+
if short_predictions > len(predictions) * 0.1: # More than 10%
|
92 |
+
warnings.append(f"{short_predictions} very short predictions (< 3 characters)")
|
93 |
+
|
94 |
+
# Check for suspiciously long predictions
|
95 |
+
long_predictions = (predictions['prediction'].str.len() > 500).sum()
|
96 |
+
if long_predictions > 0:
|
97 |
+
warnings.append(f"{long_predictions} very long predictions (> 500 characters)")
|
98 |
+
|
99 |
+
# Check for repeated predictions
|
100 |
+
duplicate_predictions = predictions['prediction'].duplicated().sum()
|
101 |
+
if duplicate_predictions > len(predictions) * 0.5: # More than 50%
|
102 |
+
warnings.append(f"{duplicate_predictions} duplicate prediction texts")
|
103 |
+
|
104 |
+
# Check for non-text content
|
105 |
+
non_text_pattern = r'^[A-Za-z\s\'".,!?;:()\-]+$'
|
106 |
+
non_text_predictions = ~predictions['prediction'].str.match(non_text_pattern, na=False)
|
107 |
+
if non_text_predictions.sum() > 0:
|
108 |
+
warnings.append(f"{non_text_predictions.sum()} predictions contain unusual characters")
|
109 |
+
|
110 |
+
return {
|
111 |
+
'has_issues': len(issues) > 0,
|
112 |
+
'issues': issues,
|
113 |
+
'warnings': warnings,
|
114 |
+
'quality_score': max(0, 1.0 - len(issues) * 0.2 - len(warnings) * 0.1)
|
115 |
+
}
|
116 |
+
|
117 |
+
def validate_against_test_set(predictions: pd.DataFrame, test_set: pd.DataFrame) -> Dict:
|
118 |
+
"""Validate predictions against the official test set."""
|
119 |
+
|
120 |
+
# Convert IDs to string for comparison
|
121 |
+
pred_ids = set(predictions['sample_id'].astype(str))
|
122 |
+
test_ids = set(test_set['sample_id'].astype(str))
|
123 |
+
|
124 |
+
# Check coverage
|
125 |
+
missing_ids = test_ids - pred_ids
|
126 |
+
extra_ids = pred_ids - test_ids
|
127 |
+
matching_ids = pred_ids & test_ids
|
128 |
+
|
129 |
+
coverage = len(matching_ids) / len(test_ids)
|
130 |
+
|
131 |
+
# Detailed coverage by language pair
|
132 |
+
pair_coverage = {}
|
133 |
+
for _, row in test_set.iterrows():
|
134 |
+
pair_key = f"{row['source_language']}_{row['target_language']}"
|
135 |
+
if pair_key not in pair_coverage:
|
136 |
+
pair_coverage[pair_key] = {'total': 0, 'covered': 0}
|
137 |
+
|
138 |
+
pair_coverage[pair_key]['total'] += 1
|
139 |
+
if str(row['sample_id']) in pred_ids:
|
140 |
+
pair_coverage[pair_key]['covered'] += 1
|
141 |
+
|
142 |
+
# Calculate pair-wise coverage rates
|
143 |
+
for pair_key in pair_coverage:
|
144 |
+
pair_info = pair_coverage[pair_key]
|
145 |
+
pair_info['coverage_rate'] = pair_info['covered'] / pair_info['total']
|
146 |
+
|
147 |
+
return {
|
148 |
+
'overall_coverage': coverage,
|
149 |
+
'missing_count': len(missing_ids),
|
150 |
+
'extra_count': len(extra_ids),
|
151 |
+
'matching_count': len(matching_ids),
|
152 |
+
'is_complete': coverage == 1.0,
|
153 |
+
'pair_coverage': pair_coverage,
|
154 |
+
'missing_ids_sample': list(missing_ids)[:10], # First 10 for display
|
155 |
+
'extra_ids_sample': list(extra_ids)[:10]
|
156 |
+
}
|
157 |
+
|
158 |
+
def generate_validation_report(
|
159 |
+
format_result: Dict,
|
160 |
+
content_result: Dict,
|
161 |
+
test_set_result: Dict,
|
162 |
+
model_name: str = ""
|
163 |
+
) -> str:
|
164 |
+
"""Generate human-readable validation report."""
|
165 |
+
|
166 |
+
report = []
|
167 |
+
|
168 |
+
# Header
|
169 |
+
report.append(f"# Validation Report: {model_name or 'Submission'}")
|
170 |
+
report.append(f"Generated: {pd.Timestamp.now().strftime('%Y-%m-%d %H:%M:%S')}")
|
171 |
+
report.append("")
|
172 |
+
|
173 |
+
# File format validation
|
174 |
+
if format_result['valid']:
|
175 |
+
report.append("β
**File Format**: Valid")
|
176 |
+
report.append(f" - Rows: {format_result['row_count']:,}")
|
177 |
+
report.append(f" - Columns: {', '.join(format_result['columns'])}")
|
178 |
+
else:
|
179 |
+
report.append("β **File Format**: Invalid")
|
180 |
+
report.append(f" - Error: {format_result['error']}")
|
181 |
+
return "\n".join(report)
|
182 |
+
|
183 |
+
# Content validation
|
184 |
+
if content_result['has_issues']:
|
185 |
+
report.append("β οΈ **Content Quality**: Issues Found")
|
186 |
+
for issue in content_result['issues']:
|
187 |
+
report.append(f" - β {issue}")
|
188 |
+
else:
|
189 |
+
report.append("β
**Content Quality**: Good")
|
190 |
+
|
191 |
+
if content_result['warnings']:
|
192 |
+
for warning in content_result['warnings']:
|
193 |
+
report.append(f" - β οΈ {warning}")
|
194 |
+
|
195 |
+
# Test set validation
|
196 |
+
coverage = test_set_result['overall_coverage']
|
197 |
+
if coverage == 1.0:
|
198 |
+
report.append("β
**Test Set Coverage**: Complete")
|
199 |
+
elif coverage >= 0.95:
|
200 |
+
report.append("β οΈ **Test Set Coverage**: Nearly Complete")
|
201 |
+
else:
|
202 |
+
report.append("β **Test Set Coverage**: Incomplete")
|
203 |
+
|
204 |
+
report.append(f" - Coverage: {coverage:.1%} ({test_set_result['matching_count']:,} / {test_set_result['matching_count'] + test_set_result['missing_count']:,})")
|
205 |
+
|
206 |
+
if test_set_result['missing_count'] > 0:
|
207 |
+
report.append(f" - Missing: {test_set_result['missing_count']:,} samples")
|
208 |
+
|
209 |
+
if test_set_result['extra_count'] > 0:
|
210 |
+
report.append(f" - Extra: {test_set_result['extra_count']:,} samples")
|
211 |
+
|
212 |
+
# Language pair coverage
|
213 |
+
pair_cov = test_set_result['pair_coverage']
|
214 |
+
incomplete_pairs = [k for k, v in pair_cov.items() if v['coverage_rate'] < 1.0]
|
215 |
+
|
216 |
+
if incomplete_pairs:
|
217 |
+
report.append("")
|
218 |
+
report.append("**Incomplete Language Pairs:**")
|
219 |
+
for pair in incomplete_pairs[:5]: # Show first 5
|
220 |
+
info = pair_cov[pair]
|
221 |
+
src, tgt = pair.split('_')
|
222 |
+
report.append(f" - {src}β{tgt}: {info['covered']}/{info['total']} ({info['coverage_rate']:.1%})")
|
223 |
+
|
224 |
+
if len(incomplete_pairs) > 5:
|
225 |
+
report.append(f" - ... and {len(incomplete_pairs) - 5} more pairs")
|
226 |
+
|
227 |
+
# Final verdict
|
228 |
+
report.append("")
|
229 |
+
if format_result['valid'] and coverage >= 0.95 and not content_result['has_issues']:
|
230 |
+
report.append("π **Overall**: Ready for evaluation!")
|
231 |
+
elif format_result['valid'] and coverage >= 0.8:
|
232 |
+
report.append("β οΈ **Overall**: Can be evaluated with warnings")
|
233 |
+
else:
|
234 |
+
report.append("β **Overall**: Please fix issues before submission")
|
235 |
+
|
236 |
+
return "\n".join(report)
|
237 |
+
|
238 |
+
def validate_submission_complete(file_content: bytes, filename: str, test_set: pd.DataFrame, model_name: str = "") -> Dict:
|
239 |
+
"""Complete validation pipeline for a submission."""
|
240 |
+
|
241 |
+
# Step 1: File format validation
|
242 |
+
format_result = validate_file_format(file_content, filename)
|
243 |
+
if not format_result['valid']:
|
244 |
+
return {
|
245 |
+
'valid': False,
|
246 |
+
'report': generate_validation_report(format_result, {}, {}, model_name),
|
247 |
+
'predictions': None
|
248 |
+
}
|
249 |
+
|
250 |
+
predictions = format_result['dataframe']
|
251 |
+
|
252 |
+
# Step 2: Content validation
|
253 |
+
content_result = validate_predictions_content(predictions)
|
254 |
+
|
255 |
+
# Step 3: Test set validation
|
256 |
+
test_set_result = validate_against_test_set(predictions, test_set)
|
257 |
+
|
258 |
+
# Step 4: Generate report
|
259 |
+
report = generate_validation_report(format_result, content_result, test_set_result, model_name)
|
260 |
+
|
261 |
+
# Overall validity
|
262 |
+
is_valid = (
|
263 |
+
format_result['valid'] and
|
264 |
+
not content_result['has_issues'] and
|
265 |
+
test_set_result['overall_coverage'] >= 0.95
|
266 |
+
)
|
267 |
+
|
268 |
+
return {
|
269 |
+
'valid': is_valid,
|
270 |
+
'coverage': test_set_result['overall_coverage'],
|
271 |
+
'report': report,
|
272 |
+
'predictions': predictions,
|
273 |
+
'pair_coverage': test_set_result['pair_coverage']
|
274 |
+
}
|