GPT Optimization of function and critical fix
This commit is contained in:
parent
7395e8aca6
commit
cfa9f6d929
1 changed files with 90 additions and 120 deletions
|
|
@ -41,6 +41,13 @@ elif 'gpt-4' in MODEL:
|
|||
INPUTAPICOST = .01
|
||||
OUTPUTAPICOST = .03
|
||||
|
||||
# Batch Size - GPT 3.5 Struggles past 15 lines per request. GPT4 struggles past 50 lines per request
|
||||
# If you are getting a MISMATCH LENGTH error, lower the batch size.
|
||||
if 'gpt-3.5' in MODEL:
|
||||
BATCHSIZE = 15
|
||||
elif 'gpt-4' in MODEL:
|
||||
BATCHSIZE = 50
|
||||
|
||||
#tqdm Globals
|
||||
BAR_FORMAT='{l_bar}{bar:10}{r_bar}{bar:-10b}'
|
||||
POSITION = 0
|
||||
|
|
@ -743,6 +750,9 @@ def searchCodes(page, pbar, fillList):
|
|||
elif finalJAString != '':
|
||||
docList.append(f'{speaker}: {finalJAString}')
|
||||
textHistory.append(finalJAString)
|
||||
else:
|
||||
docList.append(speaker)
|
||||
textHistory.append(speaker)
|
||||
speaker = ''
|
||||
match = []
|
||||
currentGroup = []
|
||||
|
|
@ -1502,6 +1512,8 @@ def searchCodes(page, pbar, fillList):
|
|||
if len(fillList) != len(docList):
|
||||
pbar.write('WARNING. LENGTH MISMATCH')
|
||||
else:
|
||||
if '氷の洞窟に' in str(docList):
|
||||
print('t')
|
||||
docList = []
|
||||
searchCodes(page, pbar, fillList)
|
||||
|
||||
|
|
@ -1822,155 +1834,113 @@ def resubVars(translatedText, allList):
|
|||
|
||||
return translatedText
|
||||
|
||||
def batch_list(input_list, batch_size):
|
||||
def batchList(input_list, batch_size):
|
||||
if not isinstance(batch_size, int) or batch_size <= 0:
|
||||
raise ValueError("batch_size must be a positive integer")
|
||||
|
||||
return [input_list[i:i + batch_size] for i in range(0, len(input_list), batch_size)]
|
||||
|
||||
def createContext(fullPromptFlag, subbedT):
|
||||
system = PROMPT if fullPromptFlag else \
|
||||
f'Output ONLY the {LANGUAGE} translation in the following format: `Translation: <{LANGUAGE.upper()}_TRANSLATION>`'
|
||||
user = f'Line to Translate = {subbedT}'
|
||||
return system, user
|
||||
|
||||
def translateText(subbedT, history, fullPromptFlag):
|
||||
system, user = createContext(fullPromptFlag, subbedT)
|
||||
# Creating message list
|
||||
msg = [{"role": "system", "content": system}]
|
||||
if isinstance(history, list):
|
||||
msg.extend([{"role": "user", "content": h} for h in history])
|
||||
else:
|
||||
msg.append({"role": "user", "content": history})
|
||||
msg.append({"role": "user", "content": user})
|
||||
response = openai.ChatCompletion.create(
|
||||
temperature=0,
|
||||
frequency_penalty=0,
|
||||
presence_penalty=0,
|
||||
model=MODEL,
|
||||
messages=msg,
|
||||
request_timeout=TIMEOUT,
|
||||
)
|
||||
return response
|
||||
|
||||
def cleanTranslatedText(translatedText, varResponse):
|
||||
placeholders = {
|
||||
f'{LANGUAGE} Translation: ': '',
|
||||
'Translation: ': '',
|
||||
# Add more replacements as needed
|
||||
}
|
||||
for target, replacement in placeholders.items():
|
||||
translatedText = translatedText.replace(target, replacement)
|
||||
|
||||
translatedText = resubVars(translatedText, varResponse[1])
|
||||
return [line for line in translatedText.split('\n') if line]
|
||||
|
||||
def extractTranslation(translatedTextList, is_list):
|
||||
pattern = r'L(\d+) - (.+)'
|
||||
# If it's a batch (i.e., list), extract with tags; otherwise, return the single item.
|
||||
if is_list:
|
||||
return [re.findall(pattern, line)[0][1] for line in translatedTextList if re.search(pattern, line)]
|
||||
else:
|
||||
matchList = re.findall(pattern, translatedTextList)
|
||||
return matchList[0][1] if matchList else translatedTextList
|
||||
|
||||
def countTokens(tItem, history):
|
||||
enc = tiktoken.encoding_for_model(MODEL)
|
||||
encode_count = lambda item: sum(len(enc.encode(line)) for line in (item if isinstance(item, list) else [item]))
|
||||
inputTotalTokens = encode_count(history) + encode_count(PROMPT)
|
||||
outputTotalTokens = encode_count(tItem) * 2 # Estimated
|
||||
return inputTotalTokens + outputTotalTokens
|
||||
|
||||
def combineList(tlist, text):
|
||||
if isinstance(text, list):
|
||||
return [t for sublist in tlist for t in sublist]
|
||||
return tlist[0]
|
||||
|
||||
@retry(exceptions=Exception, tries=5, delay=5)
|
||||
def translateGPT(text, history, fullPromptFlag):
|
||||
totalTokens = [0,0]
|
||||
# Batch
|
||||
totalTokens = [0, 0]
|
||||
if isinstance(text, list):
|
||||
tList = batch_list(text, 30)
|
||||
history = '' # We don't need history for lists
|
||||
tList = batchList(text, BATCHSIZE)
|
||||
history = ''
|
||||
else:
|
||||
tList = [text]
|
||||
|
||||
# Translate
|
||||
for index in range(len(tList)):
|
||||
tItem = tList[index]
|
||||
for index, tItem in enumerate(tList):
|
||||
# Before sending to translation, if we have a list of items, add the formatting
|
||||
if isinstance(tItem, list):
|
||||
# Combine List into payload
|
||||
payload = ''
|
||||
for i in range(len(tItem)):
|
||||
payload = payload + f'L{i} - {tItem[i]}\n'
|
||||
|
||||
payload = '\n'.join([f'L{i} - {item}' for i, item in enumerate(tItem)])
|
||||
varResponse = subVars(payload)
|
||||
subbedT = varResponse[0]
|
||||
|
||||
else:
|
||||
# Sub Vars
|
||||
varResponse = subVars(tItem)
|
||||
subbedT = varResponse[0]
|
||||
|
||||
# If there isn't any Japanese in the text just skip
|
||||
if not re.search(r'[一-龠]+|[ぁ-ゔ]+|[ァ-ヴ]+|[\uFF00-\uFFEF]', subbedT):
|
||||
tList[index] = tItem
|
||||
# Things to Check before starting translation
|
||||
if not re.search(r'[一-龠ぁ-ゔァ-ヴーａ-ｚＡ-Ｚ０-９]+', subbedT):
|
||||
continue
|
||||
|
||||
# If ESTIMATE is True just count this as an execution and return.
|
||||
if ESTIMATE:
|
||||
enc = tiktoken.encoding_for_model(MODEL)
|
||||
textRaw = ''
|
||||
historyRaw = ''
|
||||
|
||||
# Check Lists
|
||||
if isinstance(tItem, list):
|
||||
for line in tItem:
|
||||
textRaw += line
|
||||
for line in history:
|
||||
historyRaw += line
|
||||
else:
|
||||
textRaw = tItem
|
||||
historyRaw = history
|
||||
|
||||
# Calculate Estimate
|
||||
inputTotalTokens = len(enc.encode(historyRaw)) + len(enc.encode(PROMPT))
|
||||
outputTotalTokens = len(enc.encode(textRaw)) * 2 # Estimating 2x the size of the original text
|
||||
totalTokens += inputTotalTokens
|
||||
totalTokens += outputTotalTokens
|
||||
tList[index] = tItem
|
||||
totalTokens[0] += countTokens(tItem, history)
|
||||
continue
|
||||
|
||||
# Characters
|
||||
context = 'Game Characters:\
|
||||
Character: フローラ == Flora - Gender: Female\
|
||||
Character: リーゼ == Liese - Gender: Female'
|
||||
|
||||
# Prompt
|
||||
if fullPromptFlag:
|
||||
system = PROMPT
|
||||
user = 'Line to Translate = ' + subbedT
|
||||
else:
|
||||
system = 'Output ONLY the '+ LANGUAGE +' translation in the following format: `Translation: <'+ LANGUAGE.upper() +'_TRANSLATION>`'
|
||||
user = 'Line to Translate = ' + subbedT
|
||||
|
||||
# Create Message List
|
||||
msg = []
|
||||
msg.append({"role": "system", "content": system})
|
||||
msg.append({"role": "user", "content": context})
|
||||
if isinstance(history, list):
|
||||
for line in history:
|
||||
msg.append({"role": "user", "content": line})
|
||||
else:
|
||||
msg.append({"role": "user", "content": history})
|
||||
msg.append({"role": "user", "content": user})
|
||||
|
||||
response = openai.ChatCompletion.create(
|
||||
temperature=0,
|
||||
frequency_penalty=0,
|
||||
presence_penalty=0,
|
||||
model=MODEL,
|
||||
messages=msg,
|
||||
request_timeout=TIMEOUT,
|
||||
)
|
||||
|
||||
# Save Translated Text
|
||||
# Translating
|
||||
response = translateText(subbedT, history, fullPromptFlag)
|
||||
translatedText = response.choices[0].message.content
|
||||
totalTokens[0] += response.usage.prompt_tokens
|
||||
totalTokens[1] += response.usage.completion_tokens
|
||||
|
||||
# Resub Vars
|
||||
translatedText = resubVars(translatedText, varResponse[1])
|
||||
|
||||
# Remove Placeholder Text
|
||||
translatedText = translatedText.replace(LANGUAGE +' Translation: ', '')
|
||||
translatedText = translatedText.replace('Translation: ', '')
|
||||
translatedText = translatedText.replace('Line to Translate = ', '')
|
||||
translatedText = translatedText.replace('Translation = ', '')
|
||||
translatedText = translatedText.replace('Translate = ', '')
|
||||
translatedText = translatedText.replace(LANGUAGE +' Translation:', '')
|
||||
translatedText = translatedText.replace('Translation:', '')
|
||||
translatedText = translatedText.replace('Line to Translate =', '')
|
||||
translatedText = translatedText.replace('Translation =', '')
|
||||
translatedText = translatedText.replace('Translate =', '')
|
||||
translatedText = translatedText.replace('っ', '')
|
||||
translatedText = translatedText.replace('ッ', '')
|
||||
translatedText = translatedText.replace('ぁ', '')
|
||||
translatedText = translatedText.replace('～', '~')
|
||||
translatedText = translatedText.replace('。', '.')
|
||||
translatedText = translatedText.replace('、', ',')
|
||||
translatedText = translatedText.replace('？', '?')
|
||||
translatedText = translatedText.replace('！', '!')
|
||||
translatedText = translatedText.replace('「', '\"')
|
||||
translatedText = translatedText.replace('」', '\"')
|
||||
|
||||
translatedTextList = translatedText.split('\n')
|
||||
translatedTextList = list(filter(None, translatedTextList))
|
||||
|
||||
# Return Translation
|
||||
# Formatting
|
||||
translatedTextList = cleanTranslatedText(translatedText, varResponse)
|
||||
if isinstance(tItem, list):
|
||||
for i in range(len(translatedTextList)):
|
||||
matchList = re.findall(r'.*L[0-9]+\s?-\s?(.+)', translatedTextList[i])
|
||||
if len(matchList) > 0:
|
||||
translatedTextList[i] = matchList[0]
|
||||
tList[index] = translatedTextList
|
||||
history = translatedTextList[-10:] # Going to need this for the next batch
|
||||
extractedTranslations = extractTranslation(translatedTextList, True)
|
||||
tList[index] = extractedTranslations
|
||||
history = extractedTranslations[-10:] # Update history if we have a list
|
||||
else:
|
||||
matchList = re.findall(r'.*L[0-9]+\s?-\s?(.+)', translatedText)
|
||||
if len(matchList) > 0:
|
||||
translatedText = matchList[0]
|
||||
tList[index] = translatedText
|
||||
|
||||
# Combine Lists
|
||||
if isinstance(text, list):
|
||||
finalList = []
|
||||
for l in tList:
|
||||
finalList = finalList + l
|
||||
else:
|
||||
finalList = tList[0]
|
||||
# Ensure we're passing a single string to extractTranslation
|
||||
extractedTranslations = extractTranslation('\n'.join(translatedTextList), False)
|
||||
tList[index] = extractedTranslations
|
||||
|
||||
finalList = combineList(tList, text)
|
||||
return [finalList, totalTokens]
|
||||
|
|
|
|||
Loading…
Reference in a new issue