Skip to content

Commit

Permalink
Added a format based on Huggingface format (#988)
Browse files Browse the repository at this point in the history 
* Added a format based on Huggingface format

Signed-off-by: Yoav Katz <[email protected]>

* Simplified format generation (no need to tokenize)

Signed-off-by: Yoav Katz <[email protected]>

---------

Signed-off-by: Yoav Katz <[email protected]>
  • Loading branch information
@yoavkatz
yoavkatz authored Jul 4, 2024
1 parent d572545 commit 2c3b774
Show file tree
Hide file tree
Showing 2 changed files with 162 additions and 24 deletions.
132 changes: 109 additions & 23 deletions src/unitxt/formats.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,22 @@ def apply_capital_new_line_notation(text: str) -> str:
return re.sub(r"[\n(\\N)]*(\\N)+", r"\n", text)


class SystemFormat(Format):
class BaseFormat(Format):
demos_field: str = "demos"

@staticmethod
def _retrieve_field_and_pop_from_instance(instance, field_name) -> str:
if field_name is not None and field_name in instance:
field_value = instance[field_name]
instance.pop(field_name)
assert (
field_value is not None
), f"Value in field '{field_name}' should not be none. Received instance: {instance}"
return field_value
return ""


class SystemFormat(BaseFormat):
r"""Generates the whole input to the model, from constant strings that are given as args, and from values found in specified fields of the instance.
Important: formats can use '\N' notations that means new-line if no new-line before and no empty string before.
Expand Down Expand Up @@ -113,50 +128,32 @@ class SystemFormat(Format):
"""

demos_field: str = "demos"
demo_format: str = "{source}\\N{target_prefix}{target}\n\n" # example: "User: {source}\nAgent: {target}\n\n"
model_input_format: str = (
"{system_prompt}\\N{instruction}\\N{demos}{source}\\N{target_prefix}"
)
format_args: Dict[str, str] = OptionalField(default_factory=dict)

@staticmethod
def _retrieve_field_and_assert_not_none(instance, field_name) -> str:
if field_name is not None and field_name in instance:
field_value = instance[field_name]
assert (
field_value is not None
), f"Value in field '{field_name}' should not be none. Received instance: {instance}"
return field_value
return ""

def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
assert (
"source" in instance
), f"field 'source' is expected to be in the input instance. Received instance: {instance}"
source = self._retrieve_field_and_assert_not_none(
source = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="source"
)

instruction = self._retrieve_field_and_assert_not_none(
instruction = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="instruction"
)
target_prefix = self._retrieve_field_and_assert_not_none(
target_prefix = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="target_prefix"
)
system_prompt = self._retrieve_field_and_assert_not_none(
system_prompt = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="system_prompt"
)

if "target_prefix" in instance:
instance.pop("target_prefix")
if "instruction" in instance:
instance.pop("instruction")
if "system_prompt" in instance:
instance.pop("system_prompt")

demo_instances = []
if self.demos_field is not None and self.demos_field in instance:
demos = instance[self.demos_field]
Expand Down Expand Up @@ -187,3 +184,92 @@ def process(
output = apply_capital_new_line_notation(output)
instance["source"] = output
return instance


class HFSystemFormat(BaseFormat):
r"""Formats the complete input for the model using the Hugginface chat template of a given model.
HFSystemFormat expects the input instance to contain:
1. A field named "system_prompt" whose value is a string (potentially empty) that delivers a task independent opening text.
2. A field named "source" whose value is a string verbalizing the original values in the instance (as read
from the source dataset), in the context of the underlying task.
3. A field named "instruction" that contains a (non-None) string.
4. A field named with the value in arg 'demos_field', containing a list of dicts, each dict with fields "source"
and "target", representing a single demo.
5. A field named "target_prefx" that contains a string to prefix the target in both each demo, and to end the whole generated prompt
SystemFormat formats the above fields into a single string to be inputted to the model. This string overwrites
field "source" of the instance.
Example:
HFSystemFormat(model_name="HuggingFaceH4/zephyr-7b-beta")
Uses the template defined the in tokenizer_config.json of the model:
"chat_template": "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n' + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}",
See more details in https://huggingface.co/docs/transformers/main/en/chat_templating
"""

model_name: str

def process(
self, instance: Dict[str, Any], stream_name: Optional[str] = None
) -> Dict[str, Any]:
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(self.model_name)

assert (
"source" in instance
), f"field 'source' is expected to be in the input instance. Received instance: {instance}"

source = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="source"
)

instruction = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="instruction"
)
target_prefix = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="target_prefix"
)
system_prompt = self._retrieve_field_and_pop_from_instance(
instance=instance, field_name="system_prompt"
)

messages = [
{
"role": "system",
"content": system_prompt
+ ("\n" if system_prompt != "" else "")
+ instruction,
},
]
demo_instances = []
if self.demos_field is not None and self.demos_field in instance:
demos = instance[self.demos_field]
assert (
demos is not None and isoftype(demos, List[Dict[str, Any]])
), f"A list of dict-s is expected in field '{self.demos_field}'. Received instance: {instance}"
demo_instances = demos
instance.pop(self.demos_field)

for demo_instance in demo_instances:
messages.extend(
[
{"role": "user", "content": demo_instance["source"]},
{
"role": "assistant",
"content": target_prefix + demo_instance["target"],
},
]
)
messages.extend([{"role": "user", "content": source}])
tokenized_chat = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True
)

instance["source"] = tokenized_chat + target_prefix
return instance
54 changes: 53 additions & 1 deletion tests/library/test_formats.py
View file
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
from unitxt.formats import SystemFormat
from unitxt.formats import HFSystemFormat, SystemFormat
from unitxt.test_utils.operators import (
check_operator,
)
Expand All @@ -7,6 +7,58 @@


class TestFormats(UnitxtTestCase):
def test_hf_system_format(self):
instruction = "solve the math exercises"

demo_instances = [
{"source": "1+2", "target": "3", "instruction": instruction, "inputs": {}},
{"source": "4-2", "target": "2", "instruction": instruction, "inputs": {}},
]

inputs = [
{
"source": "1+1",
"target": "2",
"instruction": instruction,
"demos": demo_instances,
"inputs": {},
"target_prefix": "The answer is ",
"system_prompt": "You are a smart assistant.",
},
{
"source": "3+2",
"target": "5",
"instruction": instruction,
"demos": demo_instances,
"inputs": {},
"target_prefix": "The answer is ",
"system_prompt": "You are a smart assistant.",
},
]

# imitating iclformat's add_instruction_after_demos=True, instruction is not "", and target_prefix =""
system_format = HFSystemFormat(model_name="HuggingFaceH4/zephyr-7b-beta")

targets = [
{
"target": "2",
"inputs": {},
"source": "<|system|>\nYou are a smart assistant.\nsolve the math exercises</s>\n<|user|>\n1+2</s>\n<|assistant|>\nThe answer is 3</s>\n<|user|>\n4-2</s>\n<|assistant|>\nThe answer is 2</s>\n<|user|>\n1+1</s>\n<|assistant|>\nThe answer is ",
},
{
"target": "5",
"inputs": {},
"source": "<|system|>\nYou are a smart assistant.\nsolve the math exercises</s>\n<|user|>\n1+2</s>\n<|assistant|>\nThe answer is 3</s>\n<|user|>\n4-2</s>\n<|assistant|>\nThe answer is 2</s>\n<|user|>\n3+2</s>\n<|assistant|>\nThe answer is ",
},
]

check_operator(
operator=system_format,
inputs=inputs,
targets=targets,
tester=self,
)

def test_system_format(self):
instruction = "solve the math exercises"

Expand Down

0 comments on commit 2c3b774

Please sign in to comment.