tests/test_prompter.py

import time
import pytest

from tests.utils import wrap_test_forked
from src.enums import source_prefix, source_postfix
from src.prompter import generate_prompt

example_data_point0 = dict(instruction="Summarize",
                           input="Ducks eat seeds by the lake, then swim in the lake where fish eat small animals.",
                           output="Ducks eat and swim at the lake.")

example_data_point1 = dict(instruction="Who is smarter, Einstein or Newton?",
                           output="Einstein.")

example_data_point2 = dict(input="Who is smarter, Einstein or Newton?",
                           output="Einstein.")

example_data_points = [example_data_point0, example_data_point1, example_data_point2]


@wrap_test_forked
def test_train_prompt(prompt_type='instruct', data_point=0):
    example_data_point = example_data_points[data_point]
    return generate_prompt(example_data_point, prompt_type, '', False, False)


@wrap_test_forked
def test_test_prompt(prompt_type='instruct', data_point=0):
    example_data_point = example_data_points[data_point]
    example_data_point.pop('output', None)
    return generate_prompt(example_data_point, prompt_type, '', False, False)


@wrap_test_forked
def test_test_prompt2(prompt_type='human_bot', data_point=0):
    example_data_point = example_data_points[data_point]
    example_data_point.pop('output', None)
    res = generate_prompt(example_data_point, prompt_type, '', False, False)
    print(res, flush=True)
    return res


prompt_fastchat = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hello! ASSISTANT: Hi!</s>USER: How are you? ASSISTANT: I'm good</s>USER: Go to the market? ASSISTANT:"""

prompt_humanbot = """<human>: Hello!\n<bot>: Hi!\n<human>: How are you?\n<bot>: I'm good\n<human>: Go to the market?\n<bot>:"""

prompt_prompt_answer = "<|prompt|>Hello!<|endoftext|><|answer|>Hi!<|endoftext|><|prompt|>How are you?<|endoftext|><|answer|>I'm good<|endoftext|><|prompt|>Go to the market?<|endoftext|><|answer|>"

prompt_prompt_answer_openllama = "<|prompt|>Hello!</s><|answer|>Hi!</s><|prompt|>How are you?</s><|answer|>I'm good</s><|prompt|>Go to the market?</s><|answer|>"

prompt_mpt_instruct = """Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction
Hello!

### Response
Hi!

### Instruction
How are you?

### Response
I'm good

### Instruction
Go to the market?

### Response
"""

prompt_mpt_chat = """<|im_start|>system
A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.
<|im_end|><|im_start|>user
Hello!<|im_end|><|im_start|>assistant
Hi!<|im_end|><|im_start|>user
How are you?<|im_end|><|im_start|>assistant
I'm good<|im_end|><|im_start|>user
Go to the market?<|im_end|><|im_start|>assistant
"""

prompt_falcon = """User: Hello!

Assistant: Hi!

User: How are you?

Assistant: I'm good

User: Go to the market?

Assistant:"""

prompt_llama2 = """<s>[INST] Hello! [/INST] Hi! </s><s>[INST] How are you? [/INST] I'm good </s><s>[INST] Go to the market? [/INST]"""

prompt_llama2_sys = """<s>[INST] <<SYS>>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
<</SYS>>

Hello! [/INST] Hi! </s><s>[INST] How are you? [/INST] I'm good </s><s>[INST] Go to the market? [/INST]"""

prompt_llama2_pig = """<s>[INST] Who are you? [/INST] I am a big pig who loves to tell kid stories </s><s>[INST] Hello! [/INST] Hi! </s><s>[INST] How are you? [/INST] I'm good </s><s>[INST] Go to the market? [/INST]"""

# Fastsys doesn't put space above before final [/INST], I think wrong, since with context version has space.
# and llama2 code has space before it always: https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/generation.py


prompt_beluga = """### User:
Hello!

### Assistant:
Hi!

### User:
How are you?

### Assistant:
I'm good

### User:
Go to the market?

### Assistant:
"""

prompt_beluga_sys = """### System:
You are Stable Beluga, an AI that follows instructions extremely well. Help as much as you can. Remember, be safe, and don't do anything illegal.

### User:
Hello!

### Assistant:
Hi!

### User:
How are you?

### Assistant:
I'm good

### User:
Go to the market?

### Assistant:
"""

prompt_falcon180 = """User: Hello!
Falcon: Hi!
User: How are you?
Falcon: I'm good
User: Go to the market?
Falcon:"""

prompt_falcon180_sys = """System: You are an intelligent and helpful assistant.
User: Hello!
Falcon: Hi!
User: How are you?
Falcon: I'm good
User: Go to the market?
Falcon:"""

# below doesn't actually work for xin, use alternative that works
# prompt_xwin = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hello! ASSISTANT: Hi!</s>USER: How are you? ASSISTANT: I'm good</s>USER: Go to the market? ASSISTANT:"""
prompt_xwin = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Hello!\nASSISTANT: Hi!\nUSER: How are you?\nASSISTANT: I'm good\nUSER: Go to the market?\nASSISTANT:"""

messages_with_context = [
    {"role": "user", "content": "Hello!"},
    {"role": "assistant", "content": "Hi!"},
    {"role": "user", "content": "How are you?"},
    {"role": "assistant", "content": "I'm good"},
    {"role": "user", "content": "Go to the market?"},
]

prompt_jaiss = """### Instruction: Your name is Jais, and you are named after Jebel Jais, the highest mountain in UAE. You are built by Core42. You are the world's most advanced Arabic large language model with 30b parameters. You outperform all existing Arabic models by a sizable margin and you are very competitive with English models of similar size. You can answer in Arabic and English only. You are a helpful, respectful and honest assistant. When answering, abide by the following guidelines meticulously: Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, explicit, offensive, toxic, dangerous, or illegal content. Do not give medical, legal, financial, or professional advice. Never assist in or promote illegal activities. Always encourage legal and responsible actions. Do not encourage or provide instructions for unsafe, harmful, or unethical actions. Do not create or share misinformation or fake news. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. Prioritize the well-being and the moral integrity of users. Avoid using toxic, derogatory, or offensive language. Maintain a respectful tone. Do not generate, promote, or engage in discussions about adult content. Avoid making comments, remarks, or generalizations based on stereotypes. Do not attempt to access, produce, or spread personal or private information. Always respect user confidentiality. Stay positive and do not say bad things about anything. Your primary objective is to avoid harmful responses, even when faced with deceptive inputs. Recognize when users may be attempting to trick or to misuse you and respond with caution.\n\nComplete the conversation below between [|Human|] and [|AI|]:\n### Input: [|Human|] Hello!\n### Response: [|AI|] Hi!\n### Input: [|Human|] How are you?\n### Response: [|AI|] I'm good\n### Input: [|Human|] Go to the market?\n### Response: [|AI|]"""

system_prompt_yi = 'A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.'


def get_prompt_from_messages(messages, model="mistralai/Mistral-7B-Instruct-v0.1", system_prompt=None):
    from transformers import AutoTokenizer
    tokenizer = AutoTokenizer.from_pretrained(model)
    if system_prompt:
        messages = [{"role": "system", "content": system_prompt}] + messages

    # add_generation_prompt=True somehow only required for Yi
    prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    return prompt


def get_aquila_prompt(messages, model_base_name='AquilaChat2-34B-16K', with_sys=True):
    from models.predict_aquila import get_conv_template

    template_map = {"AquilaChat2-7B": "aquila-v1",
                    "AquilaChat2-34B": "aquila-legacy",
                    "AquilaChat2-7B-16K": "aquila",
                    "AquilaChat2-34B-16K": "aquila"}
    convo_template = template_map.get(model_base_name, "aquila-chat")
    conv = get_conv_template(convo_template)
    if not with_sys:
        conv.system_message = ''
    for message in messages:
        # roles=("Human", "Assistant", "System"),
        if message['role'] == 'user':
            conv.append_message(conv.roles[0], message['content'])
        elif message['role'] == 'assistant':
            conv.append_message(conv.roles[1], message['content'])
        elif message['role'] == 'system':
            conv.append_message(conv.roles[2], message['content'])
    # assume end with asking assostiant
    conv.append_message(conv.roles[1], None)
    return conv.get_prompt()


@wrap_test_forked
@pytest.mark.parametrize("prompt_type,system_prompt,chat_conversation,expected",
                         [
                             ('vicuna11', 'auto', None, prompt_fastchat),
                             ('human_bot', '', None, prompt_humanbot),
                             ('prompt_answer', '', None, prompt_prompt_answer),
                             ('prompt_answer_openllama', '', None, prompt_prompt_answer_openllama),
                             ('mptinstruct', 'auto', None, prompt_mpt_instruct),
                             ('mptchat', 'auto', None, prompt_mpt_chat),
                             ('falcon', '', None, prompt_falcon),
                             ('llama2', '', None, prompt_llama2),
                             ('llama2', 'auto', None, prompt_llama2_sys),
                             ('llama2', '', [('Who are you?', 'I am a big pig who loves to tell kid stories')],
                              prompt_llama2_pig),
                             ('beluga', '', None, prompt_beluga),
                             ('beluga', 'auto', None, prompt_beluga_sys),
                             ('falcon_chat', '', None, prompt_falcon180),
                             ('falcon_chat', 'auto', None, prompt_falcon180_sys),
                             ('mistral', '', None, get_prompt_from_messages(messages_with_context)),
                             ('zephyr', '', None, get_prompt_from_messages(messages_with_context,
                                                                           model='HuggingFaceH4/zephyr-7b-beta')),
                             ('zephyr', 'auto', None, get_prompt_from_messages(messages_with_context,
                                                                               model='HuggingFaceH4/zephyr-7b-beta',
                                                                               system_prompt='You are an AI that follows instructions extremely well and as helpful as possible.')),
                             ('zephyr', 'I am a cute pixie.', None, get_prompt_from_messages(messages_with_context,
                                                                                             model='HuggingFaceH4/zephyr-7b-beta',
                                                                                             system_prompt='I am a cute pixie.')),
                             ('xwin', 'auto', None, prompt_xwin),
                             ('aquila', '', None, get_aquila_prompt(messages_with_context, with_sys=False,
                                                                    model_base_name='AquilaChat2-34B-16K')),
                             ('aquila', 'auto', None, get_aquila_prompt(messages_with_context, with_sys=True,
                                                                        model_base_name='AquilaChat2-34B-16K')),
                             ('aquila_legacy', 'auto', None, get_aquila_prompt(messages_with_context, with_sys=True,
                                                                               model_base_name='AquilaChat2-34B')),
                             ('aquila_v1', 'auto', None, get_aquila_prompt(messages_with_context, with_sys=True,
                                                                           model_base_name='AquilaChat2-7B')),
                             ('deepseek_coder', 'auto', None, get_prompt_from_messages(messages_with_context,
                                                                                       model='deepseek-ai/deepseek-coder-33b-instruct')),
                             ('jais', 'auto', None, prompt_jaiss),
                             ('yi', 'auto', None,
                              get_prompt_from_messages(messages_with_context, model='01-ai/Yi-34B-Chat',
                                                       system_prompt=system_prompt_yi)),
                         ]
                         )
def test_prompt_with_context(prompt_type, system_prompt, chat_conversation, expected):
    prompt_dict = None  # not used unless prompt_type='custom'
    langchain_mode = 'Disabled'
    add_chat_history_to_context = True
    chat = True
    model_max_length = 2048
    memory_restriction_level = 0
    keep_sources_in_context = False
    iinput = ''
    stream_output = False
    debug = False

    from src.prompter import Prompter
    from src.gen import history_to_context

    t0 = time.time()
    history = [["Hello!", "Hi!"],
               ["How are you?", "I'm good"],
               ["Go to the market?", None]
               ]
    print("duration1: %s %s" % (prompt_type, time.time() - t0), flush=True)
    t0 = time.time()
    context = history_to_context(history,
                                 langchain_mode=langchain_mode,
                                 add_chat_history_to_context=add_chat_history_to_context,
                                 prompt_type=prompt_type,
                                 prompt_dict=prompt_dict,
                                 model_max_length=model_max_length,
                                 memory_restriction_level=memory_restriction_level,
                                 keep_sources_in_context=keep_sources_in_context,
                                 system_prompt=system_prompt,
                                 chat_conversation=chat_conversation)
    print("duration2: %s %s" % (prompt_type, time.time() - t0), flush=True)
    t0 = time.time()
    instruction = history[-1][0]

    # get prompt
    prompter = Prompter(prompt_type, prompt_dict, debug=debug, stream_output=stream_output,
                        system_prompt=system_prompt)
    # for instruction-tuned models, expect this:
    assert prompter.PreResponse
    assert prompter.PreInstruct
    assert prompter.botstr
    assert prompter.humanstr
    print("duration3: %s %s" % (prompt_type, time.time() - t0), flush=True)
    t0 = time.time()
    data_point = dict(context=context, instruction=instruction, input=iinput)
    prompt = prompter.generate_prompt(data_point)
    print(prompt)
    print("duration4: %s %s" % (prompt_type, time.time() - t0), flush=True)
    assert prompt == expected
    assert prompt.find(source_prefix) == -1


prompt_fastchat1 = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Go to the market? ASSISTANT:"""

prompt_humanbot1 = """<human>: Go to the market?\n<bot>:"""

prompt_prompt_answer1 = "<|prompt|>Go to the market?<|endoftext|><|answer|>"

prompt_prompt_answer_openllama1 = "<|prompt|>Go to the market?</s><|answer|>"

prompt_mpt_instruct1 = """Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction
Go to the market?

### Response
"""

prompt_mpt_chat1 = """<|im_start|>system
A conversation between a user and an LLM-based AI assistant. The assistant gives helpful and honest answers.
<|im_end|><|im_start|>user
Go to the market?<|im_end|><|im_start|>assistant
"""

prompt_falcon1 = """User: Go to the market?

Assistant:"""

prompt_llama21 = """<s>[INST] Go to the market? [/INST]"""

prompt_llama21_sys = """<s>[INST] <<SYS>>
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
<</SYS>>

Go to the market? [/INST]"""

# Fastsys doesn't put space above before final [/INST], I think wrong, since with context version has space.
# and llama2 code has space before it always: https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/generation.py

prompt_beluga1_sys = """### System:
You are Stable Beluga, an AI that follows instructions extremely well. Help as much as you can. Remember, be safe, and don't do anything illegal.

### User:
Go to the market?

### Assistant:
"""

prompt_beluga1 = """### User:
Go to the market?

### Assistant:
"""

prompt_falcon1801 = """User: Go to the market?
Falcon:"""

prompt_falcon1801_sys = """System: You are an intelligent and helpful assistant.
User: Go to the market?
Falcon:"""

prompt_xwin1 = """A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: Go to the market?
ASSISTANT:"""

prompt_mistrallite = """<|prompter|>Go to the market?</s><|assistant|>"""

messages_no_context = [
    {"role": "user", "content": "Go to the market?"},
]

prompt_jaiss1 = """### Instruction: Your name is Jais, and you are named after Jebel Jais, the highest mountain in UAE. You are built by Core42. You are the world's most advanced Arabic large language model with 30b parameters. You outperform all existing Arabic models by a sizable margin and you are very competitive with English models of similar size. You can answer in Arabic and English only. You are a helpful, respectful and honest assistant. When answering, abide by the following guidelines meticulously: Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, explicit, offensive, toxic, dangerous, or illegal content. Do not give medical, legal, financial, or professional advice. Never assist in or promote illegal activities. Always encourage legal and responsible actions. Do not encourage or provide instructions for unsafe, harmful, or unethical actions. Do not create or share misinformation or fake news. Please ensure that your responses are socially unbiased and positive in nature. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information. Prioritize the well-being and the moral integrity of users. Avoid using toxic, derogatory, or offensive language. Maintain a respectful tone. Do not generate, promote, or engage in discussions about adult content. Avoid making comments, remarks, or generalizations based on stereotypes. Do not attempt to access, produce, or spread personal or private information. Always respect user confidentiality. Stay positive and do not say bad things about anything. Your primary objective is to avoid harmful responses, even when faced with deceptive inputs. Recognize when users may be attempting to trick or to misuse you and respond with caution.\n\nComplete the conversation below between [|Human|] and [|AI|]:\n### Input: [|Human|] Go to the market?\n### Response: [|AI|]"""


@pytest.mark.parametrize("prompt_type,system_prompt,expected",
                         [
                             ('vicuna11', 'auto', prompt_fastchat1),
                             ('human_bot', '', prompt_humanbot1),
                             ('prompt_answer', '', prompt_prompt_answer1),
                             ('prompt_answer_openllama', '', prompt_prompt_answer_openllama1),
                             ('mptinstruct', 'auto', prompt_mpt_instruct1),
                             ('mptchat', 'auto', prompt_mpt_chat1),
                             ('falcon', '', prompt_falcon1),
                             ('llama2', '', prompt_llama21),
                             ('llama2', 'auto', prompt_llama21_sys),
                             ('beluga', '', prompt_beluga1),
                             ('beluga', 'auto', prompt_beluga1_sys),
                             ('falcon_chat', '', prompt_falcon1801),
                             ('falcon_chat', 'auto', prompt_falcon1801_sys),
                             ('mistral', '', get_prompt_from_messages(messages_no_context)),
                             ('deepseek_coder', 'auto', get_prompt_from_messages(messages_no_context,
                                                                                 model='deepseek-ai/deepseek-coder-33b-instruct')),
                             ('xwin', 'auto', prompt_xwin1),
                             ('mistrallite', '', prompt_mistrallite),
                             ('zephyr', 'auto', get_prompt_from_messages(messages_no_context,
                                                                         model='HuggingFaceH4/zephyr-7b-beta',
                                                                         system_prompt='You are an AI that follows instructions extremely well and as helpful as possible.')),
                             ('zephyr', '', get_prompt_from_messages(messages_no_context,
                                                                     model='HuggingFaceH4/zephyr-7b-beta')),
                             ('zephyr', 'I am a cute pixie.', get_prompt_from_messages(messages_no_context,
                                                                                       model='HuggingFaceH4/zephyr-7b-beta',
                                                                                       system_prompt='I am a cute pixie.')),
                             ('aquila', 'auto', get_aquila_prompt(messages_no_context, with_sys=True)),
                             ('aquila_legacy', 'auto',
                              get_aquila_prompt(messages_no_context, with_sys=True, model_base_name='AquilaChat2-34B')),
                             ('aquila_v1', 'auto',
                              get_aquila_prompt(messages_no_context, with_sys=True, model_base_name='AquilaChat2-7B')),
                             ('jais', 'auto', prompt_jaiss1),
                             ('yi', 'auto', get_prompt_from_messages(messages_no_context, model='01-ai/Yi-34B-Chat',
                                                                     system_prompt=system_prompt_yi)),
                         ]
                         )
@wrap_test_forked
def test_prompt_with_no_context(prompt_type, system_prompt, expected):
    prompt_dict = None  # not used unless prompt_type='custom'
    chat = True
    iinput = ''
    stream_output = False
    debug = False

    from src.prompter import Prompter
    context = ''
    instruction = "Go to the market?"

    # get prompt
    prompter = Prompter(prompt_type, prompt_dict, debug=debug, stream_output=stream_output,
                        system_prompt=system_prompt)
    # for instruction-tuned models, expect this:
    assert prompter.PreResponse
    assert prompter.PreInstruct
    assert prompter.botstr
    assert prompter.humanstr
    data_point = dict(context=context, instruction=instruction, input=iinput)
    prompt = prompter.generate_prompt(data_point)
    print(prompt)
    assert prompt == expected
    assert prompt.find(source_prefix) == -1


@wrap_test_forked
def test_source():
    prompt = "Who are you?%s\nFOO\n%s" % (source_prefix, source_postfix)
    assert prompt.find(source_prefix) >= 0


# https://huggingface.co/spaces/tiiuae/falcon-180b-demo/blob/main/app.py
def falcon180_format_prompt(message, history, system_prompt):
    prompt = ""
    if system_prompt:
        prompt += f"System: {system_prompt}\n"
    for user_prompt, bot_response in history:
        prompt += f"User: {user_prompt}\n"
        prompt += f"Falcon: {bot_response}\n"  # Response already contains "Falcon: "
    prompt += f"""User: {message}
Falcon:"""
    return prompt


@wrap_test_forked
def test_falcon180():
    prompt = "Who are you?"
    for system_prompt in ['', "Talk like a Pixie."]:
        history = [["Who are you?", "I am Falcon, a monster AI model."],
                   ["What can you do?", "I can do well on leaderboard but not actually 1st."]]
        formatted_prompt = falcon180_format_prompt(prompt, history, system_prompt)
        print(formatted_prompt)