Spaces:
Sleeping
Sleeping
Update control/recommendation_handler.py
Browse files- control/recommendation_handler.py +15 -152
control/recommendation_handler.py
CHANGED
@@ -33,10 +33,7 @@ import pandas as pd
|
|
33 |
import numpy as np
|
34 |
from sklearn.metrics.pairwise import cosine_similarity
|
35 |
import os
|
36 |
-
#os.environ['TRANSFORMERS_CACHE'] ="./models/allmini/cache"
|
37 |
-
import os.path
|
38 |
from sentence_transformers import SentenceTransformer
|
39 |
-
import pickle
|
40 |
|
41 |
def populate_json(json_file_path = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json',
|
42 |
existing_json_populated_file_path = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json'):
|
@@ -177,20 +174,9 @@ def recommend_prompt(
|
|
177 |
Raises:
|
178 |
Nothing.
|
179 |
"""
|
180 |
-
if
|
181 |
-
|
182 |
-
|
183 |
-
elif(model_id == 'intfloat/multilingual-e5-large'):
|
184 |
-
json_file = './prompt-sentences-main/prompt_sentences-multilingual-e5-large.json'
|
185 |
-
umap_model_file = './models/umap/intfloat/multilingual-e5-large/umap.pkl'
|
186 |
-
else: # fall back to all-minilm as default
|
187 |
-
json_file = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json'
|
188 |
-
umap_model_file = './models/umap/sentence-transformers/all-MiniLM-L6-v2/umap.pkl'
|
189 |
-
|
190 |
-
with open(umap_model_file, 'rb') as f:
|
191 |
-
umap_model = pickle.load(f)
|
192 |
-
|
193 |
-
prompt_json = json.load( open( json_file ) )
|
194 |
|
195 |
# Output initialization
|
196 |
out, out['input'], out['add'], out['remove'] = {}, {}, {}, {}
|
@@ -243,16 +229,17 @@ def recommend_prompt(
|
|
243 |
)
|
244 |
|
245 |
# Recommendation of values to remove from the current prompt
|
246 |
-
for sentence in input_sentences:
|
247 |
-
input_embedding =
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
|
|
256 |
|
257 |
for value_idx, v in enumerate(prompt_json['negative_values']):
|
258 |
# Dealing with values without prompts and making sure they have the same dimensions
|
@@ -347,128 +334,4 @@ def get_thresholds(
|
|
347 |
thresholds['remove_lower_threshold'] = round(remove_similarities_df.describe([.1]).loc['10%', 'similarity'], 1)
|
348 |
thresholds['remove_higher_threshold'] = round(remove_similarities_df.describe([.9]).loc['90%', 'similarity'], 1)
|
349 |
|
350 |
-
return thresholds
|
351 |
-
|
352 |
-
def recommend_local(prompt, prompt_json, model_id, model_path = './models/all-MiniLM-L6-v2/', add_lower_threshold = 0.3,
|
353 |
-
add_upper_threshold = 0.5, remove_lower_threshold = 0.1,
|
354 |
-
remove_upper_threshold = 0.5):
|
355 |
-
"""
|
356 |
-
Function that recommends prompts additions or removals
|
357 |
-
using a local model.
|
358 |
-
|
359 |
-
Args:
|
360 |
-
prompt: The entered prompt text.
|
361 |
-
prompt_json: Json file populated with embeddings.
|
362 |
-
model_id: Id of the local model.
|
363 |
-
model_path: Path to the local model.
|
364 |
-
|
365 |
-
Returns:
|
366 |
-
Prompt values to add or remove.
|
367 |
-
|
368 |
-
Raises:
|
369 |
-
Nothing.
|
370 |
-
"""
|
371 |
-
if(model_id == 'baai/bge-large-en-v1.5' ):
|
372 |
-
json_file = './prompt-sentences-main/prompt_sentences-bge-large-en-v1.5.json'
|
373 |
-
umap_model_file = './models/umap/intfloat/multilingual-e5-large/umap.pkl'
|
374 |
-
elif(model_id == 'intfloat/multilingual-e5-large'):
|
375 |
-
json_file = './prompt-sentences-main/prompt_sentences-multilingual-e5-large.json'
|
376 |
-
umap_model_file = './models/umap/intfloat/multilingual-e5-large/umap.pkl'
|
377 |
-
else: # fall back to all-minilm as default
|
378 |
-
json_file = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json'
|
379 |
-
umap_model_file = './models/umap/sentence-transformers/all-MiniLM-L6-v2/umap.pkl'
|
380 |
-
|
381 |
-
with open(umap_model_file, 'rb') as f:
|
382 |
-
umap_model = pickle.load(f)
|
383 |
-
|
384 |
-
prompt_json = json.load( open( json_file ) )
|
385 |
-
|
386 |
-
# Output initialization
|
387 |
-
out, out['input'], out['add'], out['remove'] = {}, {}, {}, {}
|
388 |
-
input_items, items_to_add, items_to_remove = [], [], []
|
389 |
-
|
390 |
-
# Spliting prompt into sentences
|
391 |
-
input_sentences = split_into_sentences(prompt)
|
392 |
-
|
393 |
-
# Recommendation of values to add to the current prompt
|
394 |
-
# Using only the last sentence for the add recommendation
|
395 |
-
model = SentenceTransformer(model_path)
|
396 |
-
input_embedding = model.encode(input_sentences[-1])
|
397 |
-
|
398 |
-
for v in prompt_json['positive_values']:
|
399 |
-
# Dealing with values without prompts and makinig sure they have the same dimensions
|
400 |
-
if(len(v['centroid']) == len(input_embedding)):
|
401 |
-
if(get_similarity(pd.DataFrame(input_embedding), pd.DataFrame(v['centroid'])) > add_lower_threshold):
|
402 |
-
closer_prompt = -1
|
403 |
-
for p in v['prompts']:
|
404 |
-
d_prompt = get_similarity(pd.DataFrame(input_embedding), pd.DataFrame(p['embedding']))
|
405 |
-
# The sentence_threshold is being used as a ceiling meaning that for high similarities the sentence/value might already be presente in the prompt
|
406 |
-
# So, we don't want to recommend adding something that is already there
|
407 |
-
if(d_prompt > closer_prompt and d_prompt > add_lower_threshold and d_prompt < add_upper_threshold):
|
408 |
-
closer_prompt = d_prompt
|
409 |
-
items_to_add.append({
|
410 |
-
'value': v['label'],
|
411 |
-
'prompt': p['text'],
|
412 |
-
'similarity': d_prompt,
|
413 |
-
'x': p['x'],
|
414 |
-
'y': p['y']})
|
415 |
-
out['add'] = items_to_add
|
416 |
-
|
417 |
-
# Recommendation of values to remove from the current prompt
|
418 |
-
i = 0
|
419 |
-
|
420 |
-
# Recommendation of values to remove from the current prompt
|
421 |
-
for sentence in input_sentences:
|
422 |
-
input_embedding = model.encode(sentence) # local
|
423 |
-
# Obtaining XY coords for input sentences from a UMAP model
|
424 |
-
if(len(prompt_json['negative_values'][0]['centroid']) == len(input_embedding) and sentence != ''):
|
425 |
-
embeddings_umap = umap_model.transform(np.expand_dims(pd.DataFrame(input_embedding).squeeze(), axis=0))
|
426 |
-
input_items.append({
|
427 |
-
'sentence': sentence,
|
428 |
-
'x': str(embeddings_umap[0][0]),
|
429 |
-
'y': str(embeddings_umap[0][1])
|
430 |
-
})
|
431 |
-
|
432 |
-
for v in prompt_json['negative_values']:
|
433 |
-
# Dealing with values without prompts and makinig sure they have the same dimensions
|
434 |
-
if(len(v['centroid']) == len(input_embedding)):
|
435 |
-
if(get_similarity(pd.DataFrame(input_embedding), pd.DataFrame(v['centroid'])) > remove_lower_threshold):
|
436 |
-
closer_prompt = -1
|
437 |
-
for p in v['prompts']:
|
438 |
-
d_prompt = get_similarity(pd.DataFrame(input_embedding), pd.DataFrame(p['embedding']))
|
439 |
-
# A more restrict threshold is used here to prevent false positives
|
440 |
-
# The sentence_threhold is being used to indicate that there must be a sentence in the prompt that is similiar to one of our adversarial prompts
|
441 |
-
# So, yes, we want to recommend the revolval of something adversarial we've found
|
442 |
-
if(d_prompt > closer_prompt and d_prompt > remove_upper_threshold):
|
443 |
-
closer_prompt = d_prompt
|
444 |
-
items_to_remove.append({
|
445 |
-
'value': v['label'],
|
446 |
-
'sentence': sentence,
|
447 |
-
'sentence_index': i,
|
448 |
-
'closest_harmful_sentence': p['text'],
|
449 |
-
'similarity': d_prompt,
|
450 |
-
'x': p['x'],
|
451 |
-
'y': p['y']})
|
452 |
-
out['remove'] = items_to_remove
|
453 |
-
i += 1
|
454 |
-
|
455 |
-
out['input'] = input_items
|
456 |
-
|
457 |
-
out['add'] = sorted(out['add'], key=sort_by_similarity, reverse=True)
|
458 |
-
values_map = {}
|
459 |
-
for item in out['add'][:]:
|
460 |
-
if(item['value'] in values_map):
|
461 |
-
out['add'].remove(item)
|
462 |
-
else:
|
463 |
-
values_map[item['value']] = item['similarity']
|
464 |
-
out['add'] = out['add'][0:5]
|
465 |
-
|
466 |
-
out['remove'] = sorted(out['remove'], key=sort_by_similarity, reverse=True)
|
467 |
-
values_map = {}
|
468 |
-
for item in out['remove'][:]:
|
469 |
-
if(item['value'] in values_map):
|
470 |
-
out['remove'].remove(item)
|
471 |
-
else:
|
472 |
-
values_map[item['value']] = item['similarity']
|
473 |
-
out['remove'] = out['remove'][0:5]
|
474 |
-
return out
|
|
|
33 |
import numpy as np
|
34 |
from sklearn.metrics.pairwise import cosine_similarity
|
35 |
import os
|
|
|
|
|
36 |
from sentence_transformers import SentenceTransformer
|
|
|
37 |
|
38 |
def populate_json(json_file_path = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json',
|
39 |
existing_json_populated_file_path = './prompt-sentences-main/prompt_sentences-all-minilm-l6-v2.json'):
|
|
|
174 |
Raises:
|
175 |
Nothing.
|
176 |
"""
|
177 |
+
if embedding_fn is None:
|
178 |
+
# Use all-MiniLM-L6-v2 locally by default
|
179 |
+
embedding_fn = get_embedding_func('local', model_id='sentence-transformers/all-MiniLM-L6-v2')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
|
181 |
# Output initialization
|
182 |
out, out['input'], out['add'], out['remove'] = {}, {}, {}, {}
|
|
|
229 |
)
|
230 |
|
231 |
# Recommendation of values to remove from the current prompt
|
232 |
+
for sent_idx, sentence in enumerate(input_sentences):
|
233 |
+
input_embedding = inp_sentence_embeddings[sent_idx]
|
234 |
+
if umap_model:
|
235 |
+
# Obtaining XY coords for input sentences from a parametric UMAP model
|
236 |
+
if(len(prompt_json['negative_values'][0]['centroid']) == len(input_embedding) and sentence != ''):
|
237 |
+
embeddings_umap = umap_model.transform(np.expand_dims(pd.DataFrame(input_embedding).squeeze(), axis=0))
|
238 |
+
input_items.append({
|
239 |
+
'sentence': sentence,
|
240 |
+
'x': str(embeddings_umap[0][0]),
|
241 |
+
'y': str(embeddings_umap[0][1])
|
242 |
+
})
|
243 |
|
244 |
for value_idx, v in enumerate(prompt_json['negative_values']):
|
245 |
# Dealing with values without prompts and making sure they have the same dimensions
|
|
|
334 |
thresholds['remove_lower_threshold'] = round(remove_similarities_df.describe([.1]).loc['10%', 'similarity'], 1)
|
335 |
thresholds['remove_higher_threshold'] = round(remove_similarities_df.describe([.9]).loc['90%', 'similarity'], 1)
|
336 |
|
337 |
+
return thresholds
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|