mmarch/text-generation-webui
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Add 'hallucinations' filter #326

Browse source 
This breaks the API since a new parameter has been added.
It should be a one-line fix. See api-example.py.
This commit is contained in:
oobabooga 2023-03-15 11:04:30 -03:00
parent 128d18e298
commit 9d6a625bd6
5 changed files with 25 additions and 18 deletions
Download patch file Download diff file Expand all files Collapse all files

16
modules/chat.py
View file

@ -97,7 +97,7 @@ def extract_message_from_reply(question, reply, name1, name2, check, impersonate
def stop_everything_event():
shared.stop_everything = True
def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1, regenerate=False):
def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1, regenerate=False):
shared.stop_everything = False
just_started = True
eos_token = '\n' if check else None
@ -133,7 +133,7 @@ def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical
# Generate
reply = ''
for i in range(chat_generation_attempts):
for reply in generate_reply(f"{prompt}{' ' if len(reply) > 0 else ''}{reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name1}:"):
for reply in generate_reply(f"{prompt}{' ' if len(reply) > 0 else ''}{reply}", max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name1}:"):
# Extracting the reply
reply, next_character_found = extract_message_from_reply(prompt, reply, name1, name2, check)
@ -160,7 +160,7 @@ def chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical
yield shared.history['visible']
def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
eos_token = '\n' if check else None
if 'pygmalion' in shared.model_name.lower():
@ -172,18 +172,18 @@ def impersonate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typ
# Yield *Is typing...*
yield shared.processing_message
for i in range(chat_generation_attempts):
for reply in generate_reply(prompt+reply, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name2}:"):
for reply in generate_reply(prompt+reply, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=eos_token, stopping_string=f"\n{name2}:"):
reply, next_character_found = extract_message_from_reply(prompt, reply, name1, name2, check, impersonate=True)
yield reply
if next_character_found:
break
yield reply
def cai_chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
for _history in chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts):
def cai_chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
for _history in chatbot_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts):
yield generate_chat_html(_history, name1, name2, shared.character)
def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts=1):
if (shared.character != 'None' and len(shared.history['visible']) == 1) or len(shared.history['internal']) == 0:
yield generate_chat_output(shared.history['visible'], name1, name2, shared.character)
else:
@ -191,7 +191,7 @@ def regenerate_wrapper(text, max_new_tokens, do_sample, temperature, top_p, typi
last_internal = shared.history['internal'].pop()
# Yield '*Is typing...*'
yield generate_chat_output(shared.history['visible']+[[last_visible[0], shared.processing_message]], name1, name2, shared.character)
for _history in chatbot_wrapper(last_internal[0], max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts, regenerate=True):
for _history in chatbot_wrapper(last_internal[0], max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, name1, name2, context, check, chat_prompt_size, chat_generation_attempts, regenerate=True):
if shared.args.cai_chat:
shared.history['visible'][-1] = [last_visible[0], _history[-1][1]]
else:

3
modules/text_generation.py
View file

@ -89,7 +89,7 @@ def clear_torch_cache():
if not shared.args.cpu:
torch.cuda.empty_cache()
def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=None, stopping_string=None):
def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typical_p, repetition_penalty, encoder_repetition_penalty, top_k, min_length, no_repeat_ngram_size, num_beams, penalty_alpha, length_penalty, early_stopping, eos_token=None, stopping_string=None):
clear_torch_cache()
t0 = time.time()
@ -143,6 +143,7 @@ def generate_reply(question, max_new_tokens, do_sample, temperature, top_p, typi
"top_p": top_p,
"typical_p": typical_p,
"repetition_penalty": repetition_penalty,
"encoder_repetition_penalty": encoder_repetition_penalty,
"top_k": top_k,
"min_length": min_length if shared.args.no_stream else 0,
"no_repeat_ngram_size": no_repeat_ngram_size,