generation

Classes:

Name	Description
StainedGlassGenerationConfig	A transformers.GenerationConfig with tokenizer aware settings, with empirically optimized settings for Stained Glass.

StainedGlassGenerationConfig

Bases: GenerationConfig

A transformers.GenerationConfig with tokenizer aware settings, with empirically optimized settings for Stained Glass.

Methods:

Name	Description
__init__	Create a StainedGlassGenerationConfig.
from_tokenizer	Create a StainedGlassGenerationConfig using a tokenizer.

__init__

__init__(
    max_length: int,
    bos_token_id: int | None,
    pad_token_id: int | None,
    eos_token_id: int | None,
    temperature: float = 0.6,
    top_k: int = 5000,
    top_p: float = 0.9,
    repetition_penalty: float = 1.0,
    do_sample: bool = True,
    num_return_sequences: Literal[1] = 1,
    renormalize_logits: Literal[True] = True,
    **kwargs: Any,
) -> None

Create a StainedGlassGenerationConfig.

Parameters:

Name	Type	Description	Default
max_length	int	The maximum number of tokens in the prompt and generated text combined.	required
bos_token_id	int | None	The token id for the beginning of the sequence.	required
pad_token_id	int | None	The token id for padding.	required
eos_token_id	int | None	The token id for the end of the sequence.	required
temperature	float	A setting controlling how conservative or creative the model responses are. Lower values such as 0.2 are terse, higher values such as above 1.0 are verbose and creative.	0.6
top_k	int	A positive integer representing the maximum number of tokens considered for sampling, ordered by likelihood.	5000
top_p	float	A positive real number in (0, 1.0) representing the amount of probabilistic mass which contributes to sampling tokens.	0.9
repetition_penalty	float	A penalty factor on generating repeated tokens.	1.0
do_sample	bool	Whether or not to use sampling ; use greedy decoding otherwise.	True
num_return_sequences	Literal[1]	The number of independently computed returned sequences for each element in the batch.	1
renormalize_logits	Literal[True]	Whether to renormalize the logits after applying all the logits processors or warper. It's highly recommended to set this flag to True as the search algorithms suppose the score logits are normalized.	True
**kwargs	Any	Additional keyword arguments to the generation config.	required

from_tokenizer classmethod

from_tokenizer(
    tokenizer: PreTrainedTokenizerBase,
    max_length: int,
    **kwargs: Any,
) -> StainedGlassGenerationConfig

Create a StainedGlassGenerationConfig using a tokenizer.

Parameters:

Name	Type	Description	Default
tokenizer	PreTrainedTokenizerBase	The tokenizer whose pad, bos, and eos tokens are used for generation config.	required
max_length	int	The maximum number of tokens in the prompt and generated text combined.	required
**kwargs	Any	Additional keyword arguments to the generation config.	required

Returns:

Type	Description
StainedGlassGenerationConfig	A Stained Glass generation config.