universal
Model-agnostic Mapper classes (designed to be compatible with datasets.Dataset.map) useful for building LLM prompts for Stained
Glass Transform training and testing.
dataclass
¶
Builds the tensor components of the transformers.PreTrainedModel chat prompt.
Added in version 0.77.0.
dataclass
¶
dataclass
¶
Bases: SchemaMapper
Maps samples from an arbitrary dataset to a universal schema for building an LLM chat prompt.
Either define a subclass for easier reuse, or use this directly.
Examples:
>>> sample = {
... "question": "What is the capital of France?",
... "response": "Paris",
... "system_prompt": "Answer the following question:",
... }
>>> mapper = ChatSchemaMapper(
... instruction_key="question",
... response_key="response",
... system_prompt_key="system_prompt",
... )
>>> mapped_sample = mapper(sample)
>>> mapped_sample
[{'role': 'system', 'content': 'Answer the following question:'}, {'role': 'user', 'content': 'What is the capital of France?'}, {'role': 'assistant', 'content': 'Paris'}]
Added in version 0.77.0.
instance-attribute
¶
instruction_key: str
The dataset key/column corresponding to the input.
instance-attribute
¶
response_key: str | None
An optional dataset key/column corresponding to the expected model response to the instruction.
dataclass
¶
Special string components of a chat prompt.
An instance of this class is expected to be defined for each model to dictate the structure of its prompt.
Added in version 0.77.0.
dataclass
¶
Bases: TokenizerMapper, ABC
Tokenizes and builds the intermediate tensor components of a chat prompt.
Added in version 0.77.0.
class-attribute
instance-attribute
¶
special_strings: ChatSpecialStrings = field(init=False)
The special prompt strings to use.
class-attribute
instance-attribute
¶
special_tokens: SpecialTokens = field(init=False)
The tokenized special prompt strings.
Bases: TypedDict
Collection of all tokenized components of the prompt.
instance-attribute
¶
schema_tokens: list[SchemaTokens]
The tokenized schema components of the prompt.
instance-attribute
¶
special_tokens: SpecialTokens
The tokenized special components of the prompt.
Bases: TypedDict
Tokenized special components of the prompt.
dataclass
¶
Builds the tensor components of the transformers.PreTrainedModel instruction prompt.
Bases: TypedDict
Indices of the prompt components in the input_ids tensor.
Can be used to extract the prompt components from the input_ids tensor by slicing along the sequence dimension.
Examples:
Using the PromptIndices to extract the instruction from the input_ids tensor:
>>> mapper = InstructionFormatMapper()
>>> sample: universal.InstructionTokenizerMapper.PromptTokens = {
... "special_tokens": {
... "bos": torch.tensor([[1]]),
... "instruction_start": torch.tensor([[2]]),
... "system_prompt_start": torch.tensor([[3]]),
... "system_prompt_end": torch.tensor([[4]]),
... "context_start": torch.tensor([[5]]),
... "instruction_end": torch.tensor([[6]]),
... "eos": torch.tensor([[7]]),
... },
... "schema_tokens": {
... "instruction": torch.tensor([[8, 9, 10, 11, 12]]),
... "response": torch.tensor([[13, 14, 15]]),
... "system_prompt": torch.tensor([[16, 17, 18, 19]]),
... "context": torch.tensor([[20, 21, 22]]),
... },
... }
>>> formatted_sample = mapper(sample)
>>> torch.testing.assert_close(
... sample["schema_tokens"]["instruction"],
... formatted_sample["input_ids"][:, mapper.prompt_indices["instruction"]],
... )
dataclass
¶
Bases: SchemaMapper
Maps samples from an arbitrary dataset to a universal schema for building an LLM instruction prompt.
Either define a subclass for easier reuse, or use this class directly.
Examples:
>>> sample = {
... "question": "What is the capital of France?",
... "response": "Paris",
... "system_prompt": "Answer the following question:",
... }
>>> mapper = InstructionSchemaMapper(
... instruction_key="question",
... response_key="response",
... system_prompt_key="system_prompt",
... context_key=None,
... )
>>> mapped_sample = mapper(sample)
>>> mapped_sample
{'instruction': 'What is the capital of France?', 'response': 'Paris', 'context': '', 'system_prompt': 'Answer the following question:'}
instance-attribute
¶
context_key: str | None
An optional dataset key/column corresponding to context to append to the instruction.
instance-attribute
¶
instruction_key: str
The dataset key/column corresponding to the input.
instance-attribute
¶
response_key: str | None
An optional dataset key/column corresponding to the expected model response to the instruction.
instance-attribute
¶
system_prompt_key: str | None
An optional dataset key/column corresponding to the system prompt for the model.
dataclass
¶
Special string components of an instruction-tuning prompt.
An instance of this class is expected to be defined for each model to dictate the structure of its prompt.
Added in version 0.77.0. Renamed `SpecialStrings` to `InstructionSpecialStrings`.
instance-attribute
¶
The delimiter between the instruction and the context.
instance-attribute
¶
The end of the instruction tag. The model is highly sensitive to this tag.
instance-attribute
¶
The start of the instruction. The model is highly sensitive to this tag.
dataclass
¶
Bases: TokenizerMapper, ABC
Tokenizes and builds the intermediate tensor components of an instruction prompt.
class-attribute
instance-attribute
¶
always_include_context: bool = False
Whether to always include the start of context tokens in the prompt, even if no context is provided.
class-attribute
instance-attribute
¶
special_strings: InstructionSpecialStrings = field(
init=False
)
The special prompt strings to use.
class-attribute
instance-attribute
¶
special_tokens: SpecialTokens = field(init=False)
The tokenized special prompt strings.
Bases: TypedDict
Collection of all tokenized components of the prompt.
instance-attribute
¶
schema_tokens: SchemaTokens
The tokenized schema components of the prompt.
instance-attribute
¶
special_tokens: SpecialTokens
The tokenized special components of the prompt.
Bases: TypedDict
Tokenized intermediate prompt schema.
dataclass
¶
Builds the tensor components of the transformers.PreTrainedModel pretraining prompt.
Added in version 0.77.0. Added support for pretraining which does not use a prompt template.
dataclass
¶
Bases: SchemaMapper
Maps samples from an arbitrary dataset to a universal schema for building an LLM instruction prompt.
Either define a subclass for easier reuse, or use this class directly.
Examples:
>>> sample = {
... "question": "What is the capital of France?",
... }
>>> mapper = PreTrainSchemaMapper(
... instruction_key="question",
... )
>>> mapped_sample = mapper(sample)
>>> mapped_sample
{'text': 'What is the capital of France?'}
Added in version 0.77.0. Added support for pretraining which does not use a prompt template.
dataclass
¶
Bases: TokenizerMapper
Tokenizes and builds the intermediate tensor components of a pretraining input which does not have a prompt.
Added in version 0.77.0. Added support for pretraining which does not use a prompt template.
class-attribute
instance-attribute
¶
special_tokens: SpecialTokens = field(init=False)
The tokenized special prompt strings.
Bases: TypedDict
Collection of all tokenized components of the prompt.
instance-attribute
¶
schema_tokens: SchemaTokens
The tokenized schema components of the prompt.
instance-attribute
¶
special_tokens: SpecialTokens
The tokenized special components of the prompt.
dataclass
¶
Bases: ABC
Maps samples from an arbitrary dataset to a universal schema for building an LLM prompt.
Added in version 0.77.0. Base class for `InstructionSchemaMapper` and `ChatSchemaMapper`.
dataclass
¶
Maps a dictionary of int64 tensors to a dictionary of lists of int.
dataclass
¶
Formats the undifferentiated LlamaForCausalLM input for testing.
Added in version 0.77.0. Renamed `InstructionTestMapper` to `TestMapper`.
Bases: TypedDict, Generic[ContainerT]
Input for LlamaForCausalLM testing.
dataclass
¶
Bases: ABC
Tokenizes and builds the intermediate tensor components of a prompt.
Added in version 0.77.0. Base class for `InstructionTokenizerMapper` and `ChatTokenizerMapper`.
dataclass
¶
Formats the undifferentiated transformers.PreTrainedModel input for training.
Added in version 0.77.0. Renamed `InstructionTrainMapper` to `TrainMapper`.
dataclass
¶
Bases: TransformLayerFormatMapper, ChatFormatMapper
Builds the noise token mask for a chat prompt, which is required for training a TransformLayer.
Added in version 0.77.0.
Changed in version 0.100.0: Removed the option of passing `obfuscate_system_prompt` to the TokenizerWrapper
dataclass
¶
Base class for building noise token mask.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
|
bool
|
Whether to to transform all the tokens, or only the instruction, context, and possibly the system prompt. |
False
|
Added in version 0.77.0. Base class for `TransformLayerInstructionFormatMapper` and `TransformLayerChatFormatMapper`.
Changed in version 0.100.0: Removed the option of passing `obfuscate_system_prompt` to the TokenizerWrapper
dataclass
¶
Bases: TransformLayerFormatMapper, InstructionFormatMapper
Builds the noise token mask for a instruction prompt, which is required for training a TransformLayer.
Bases: TypedDict
Indices of the prompt components in the input_ids tensor.
Can be used to extract the prompt components from the input_ids tensor by slicing along the sequence dimension.
Examples:
Using the PromptIndices to extract the instruction from the input_ids tensor:
>>> mapper = InstructionFormatMapper()
>>> sample: universal.InstructionTokenizerMapper.PromptTokens = {
... "special_tokens": {
... "bos": torch.tensor([[1]]),
... "instruction_start": torch.tensor([[2]]),
... "system_prompt_start": torch.tensor([[3]]),
... "system_prompt_end": torch.tensor([[4]]),
... "context_start": torch.tensor([[5]]),
... "instruction_end": torch.tensor([[6]]),
... "eos": torch.tensor([[7]]),
... },
... "schema_tokens": {
... "instruction": torch.tensor([[8, 9, 10, 11, 12]]),
... "response": torch.tensor([[13, 14, 15]]),
... "system_prompt": torch.tensor([[16, 17, 18, 19]]),
... "context": torch.tensor([[20, 21, 22]]),
... },
... }
>>> formatted_sample = mapper(sample)
>>> torch.testing.assert_close(
... sample["schema_tokens"]["instruction"],
... formatted_sample["input_ids"][:, mapper.prompt_indices["instruction"]],
... )
__call__(
sample: PromptTokens,
) -> UndifferentiatedTransformLayerInput
Changed in version 0.74.0: The `noise_token_mask` was renamed to `noise_mask` to create a uniform interface everywhere.
Changed in version 0.100.0: Removed the option of passing `obfuscate_system_prompt` to the TokenizerWrapper
dataclass
¶
Bases: TransformLayerFormatMapper, PreTrainFormatMapper
Builds the noise token mask for a pretraining scenario which does not use a templated prompt, which is required for training a
TransformLayer.
Added in version 0.77.0. Added support for pretraining which does not use a prompt template.
dataclass
¶
Bases: TestMapper
Formats the undifferentiated InstructionTransformLayer input for testing.
Added in version 0.77.0. Renamed `TransformLayerInstructionTestMapper` to `TransformLayerTestMapper`.
Bases: TypedDict, Generic[ContainerT]
Input for LlamaForCausalLM testing.
dataclass
¶
Bases: TrainMapper
Formats the undifferentiated InstructionTransformLayer input for training.
Added in version 0.77.0. Renamed `TransformLayerInstructionTrainMapper` to `TransformLayerTrainMapper`.
class-attribute
instance-attribute
¶
ignore_prompt_loss: bool = True
Whether to ignore the loss on the prompt tokens.
Bases: TrainInput[ContainerT]
Input for TransformLayer training.
__call__(
sample: UndifferentiatedTransformLayerInput,
) -> TransformLayerTrainInput[torch.Tensor]
Changed in version 0.74.0: The `noise_token_mask` was renamed to `noise_mask` to create a uniform interface everywhere.
Bases: TypedDict
Formatted input for the transformers.PreTrainedModel that must be further formatted into either training or testing input.
Must be further formatted based on if the model is being trained or evaluated.
Added in version 0.77.0. Renamed `InstructionFormatMapper.UndifferentiatedInstructionInput` to `UndifferentiatedInput`.
Bases: UndifferentiatedInput
Formatted input for the TransformLayer that must be further formatted into
either training or testing input.
Must be further formatted based on if the model is being trained or evaluated.
Changed in version 0.74.0: The `noise_token_mask` was renamed to `noise_mask` to create a uniform interface everywhere.
Added in version 0.77.0. Renamed `TransformLayerInstructionFormatMapper.UndifferentiatedTransformLayerInstructionInput` to `UndifferentiatedTransformLayerInput`.