def get_final_chat_data(self) -> Dict[str, Any]:
"""
Add relevant fields to the final chat data.
"""
if self.check_acceptability:
human_messages, violation_types = self._prepare_acceptability_checking(
)
violations_string = self.acceptability_checker.check_messages(
messages=human_messages,
is_worker_0=False,
violation_types=violation_types,
)
else:
violations_string = None
data = {
'dialog': self.dialog,
'workers': [get_mturk_id_from_mephisto_wrapper(self.agent)],
'bad_workers': [],
'acceptability_violations': (violations_string, ),
'hit_ids': [self.agent.mephisto_agent.task_run_id],
'assignment_ids': [self.agent.mephisto_agent.assignment_id],
'task_description': {
'annotations_config': self.opt['annotations_config'],
'model_1_nickname': self.bots[0].worker_id,
'model_1_file': self.bots[0].model_agent.opt.get('model_file'),
'model_1_opt': self.bots[0].model_agent.opt,
'model_2_nickname': self.bots[1].worker_id,
'model_2_file': self.bots[1].model_agent.opt.get('model_file'),
'model_2_opt': self.bots[1].model_agent.opt,
},
}
# 'bad_workers' is for compatibility. Before, it was only non-empty if a
# worker abandoned, returned, etc. a HIT, but now we don't even save chat
# data in that case
if self.check_acceptability:
data['acceptability_violations'] = (violations_string, )
# Make a tuple for compatibility with a human/human conversation in
# which we check both sides for acceptability
context_data = {
'personas':
self.personas,
'context_dataset':
self.context_info.get('context_dataset'),
'person1_seed_utterance':
self.context_info.get('person1_seed_utterance'),
'person2_seed_utterance':
self.context_info.get('person2_seed_utterance'),
'additional_context':
self.context_info.get('additional_context'),
}
data.update(context_data)
return data
def __init__(self, opt, agent: "MephistoAgentWrapper"):
super().__init__(opt, agent)
self.min_correct = ONBOARD_CONFIG['min_correct']
self.max_incorrect = ONBOARD_CONFIG['max_incorrect']
self.onboard_task_data = opt['onboard_task_data']
self.status = 'DISCONNECT'
self.onboard_statistics = opt['onboard_statistics']
self.statistics_condition = opt['statistics_condition']
self.max_onboard_time = opt['max_onboard_time']
self.onboarding_qualification = opt['onboarding_qualification']
self.worker_id = get_mturk_id_from_mephisto_wrapper(self.agent)
def __init__(self, opt, agent: "MephistoAgentWrapper"):
super().__init__(opt, agent)
self.skip_onboarding = opt['skip_onboarding']
self.onboard_task_data = opt['onboard_task_data']
self.status = 'DISCONNECT'
self.onboard_statistics = opt['onboard_statistics']
self.statistics_condition = opt['statistics_condition']
self.max_onboard_time = opt['max_onboard_time']
self.onboarding_qualification = opt['onboarding_qualification']
self.worker_id = get_mturk_id_from_mephisto_wrapper(self.agent)
self.annotations = None
def get_final_chat_data(self) -> Dict[str, Any]:
"""
Return specific info about the conversation, the context, acceptability, etc.
"""
if self.check_acceptability:
human_messages, violation_types = self._prepare_acceptability_checking(
)
violations_string = self.acceptability_checker.check_messages(
messages=human_messages,
is_worker_0=False,
violation_types=violation_types,
)
else:
violations_string = None
data = {
'dialog': self.dialog,
'workers': [get_mturk_id_from_mephisto_wrapper(self.agent)],
'bad_workers': [],
'acceptability_violations': (violations_string, ),
'hit_ids': [self.agent.mephisto_agent.task_run_id],
'assignment_ids': [self.agent.mephisto_agent.assignment_id],
'task_description': {
'annotations_config': self.opt['annotations_config'],
'model_nickname': self.bot.worker_id,
'model_file': self.bot.model_agent.opt.get('model_file'),
'model_opt': self.bot.model_agent.opt,
},
}
# TODO: once the analysis scripts are fully switched over to DataBrowser, remove
# the 'workers' and 'assignment_ids' keys, which will now be duplicated in the
# returned Unit
# TODO: 'bad_workers' is for compatibility. Before, it was only non-empty if a
# worker abandoned, returned, etc. a HIT, but now we don't even save chat
# data in that case. Remove this key once fully once on DataBrowser
if self.check_acceptability:
data['acceptability_violations'] = (violations_string, )
# Make a tuple for compatibility with a human/human conversation in
# which we check both sides for acceptability
return data
def get_final_chat_data(self) -> Dict[str, Any]:
"""
Return specific info about the conversation, the context, acceptability, etc.
"""
if self.check_acceptability:
human_texts = [
message['text'] for message in self.dialog
if message['agent_idx'] == 0
]
violation_types = [
'min_words', 'all_caps', 'exact_match', 'safety'
]
if self.opt['conversation_start_mode'] == 'bst':
# The BST mode starts the conversation with two previous utterances, so
# there should be no new greeting. Also, the first human response is one
# of the previous utterances, so it shouldn't get checked.
violation_types.append('penalize_greetings')
human_texts = human_texts[1:]
violations_string = self.acceptability_checker.check_messages(
messages=human_texts,
is_worker_0=False,
violation_types=violation_types)
else:
violations_string = None
data = {
'personas':
self.personas,
'context_dataset':
self.context_info.get('context_dataset'),
'person1_seed_utterance':
self.context_info.get('person1_seed_utterance'),
'person2_seed_utterance':
self.context_info.get('person2_seed_utterance'),
'additional_context':
self.context_info.get('additional_context'),
'dialog':
self.dialog,
'workers': [get_mturk_id_from_mephisto_wrapper(self.agent)],
'bad_workers': [],
'acceptability_violations': (violations_string, ),
'hit_ids': [self.agent.mephisto_agent.task_run_id],
'assignment_ids': [self.agent.mephisto_agent.assignment_id],
'task_description': {
'annotations_config': self.opt['annotations_config'],
'model_nickname': self.bot.worker_id,
'model_file': self.bot.model_agent.opt.get('model_file'),
'model_opt': self.bot.model_agent.opt,
},
}
# 'bad_workers' is for compatibility. Before, it was only non-empty if a
# worker abandoned, returned, etc. a HIT, but now we don't even save chat
# data in that case
if self.check_acceptability:
data['acceptability_violations'] = (violations_string, )
# Make a tuple for compatibility with a human/human conversation in
# which we check both sides for acceptability
return data
