Source code for sparknlp.annotator.seq2seq.mistral_transformer

#  Copyright 2017-2022 John Snow Labs
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
"""Contains classes for the MistralTransformer."""

from sparknlp.common import *


[docs]class MistralTransformer(AnnotatorModel, HasBatchedAnnotate, HasEngine): """Mistral 7B Mistral 7B, a 7.3 billion-parameter model that stands out for its efficient and effective performance in natural language processing. Surpassing Llama 2 13B across all benchmarks and excelling over Llama 1 34B in various aspects, Mistral 7B strikes a balance between English language tasks and code comprehension, rivaling the capabilities of CodeLlama 7B in the latter. Mistral 7B introduces Grouped-query attention (GQA) for quicker inference, enhancing processing speed without compromising accuracy. This streamlined approach ensures a smoother user experience, making Mistral 7B a practical choice for real-world applications. Additionally, Mistral 7B adopts Sliding Window Attention (SWA) to efficiently handle longer sequences at a reduced computational cost. This feature enhances the model's ability to process extensive textual input, expanding its utility in handling more complex tasks. In summary, Mistral 7B represents a notable advancement in language models, offering a reliable and versatile solution for various natural language processing challenges. Pretrained models can be loaded with :meth:`.pretrained` of the companion object: >>> mistral = MistralTransformer.pretrained() \\ ... .setInputCols(["document"]) \\ ... .setOutputCol("generation") The default model is ``"mistral_7b"``, if no name is provided. For available pretrained models please see the `Models Hub <https://sparknlp.org/models?q=mistral>`__. ====================== ====================== Input Annotation types Output Annotation type ====================== ====================== ``DOCUMENT`` ``DOCUMENT`` ====================== ====================== Parameters ---------- configProtoBytes ConfigProto from tensorflow, serialized into byte array. minOutputLength Minimum length of the sequence to be generated, by default 0 maxOutputLength Maximum length of output text, by default 20 doSample Whether or not to use sampling; use greedy decoding otherwise, by default False temperature The value used to module the next token probabilities, by default 1.0 topK The number of highest probability vocabulary tokens to keep for top-k-filtering, by default 50 topP Top cumulative probability for vocabulary tokens, by default 1.0 If set to float < 1, only the most probable tokens with probabilities that add up to ``topP`` or higher are kept for generation. repetitionPenalty The parameter for repetition penalty, 1.0 means no penalty. , by default 1.0 noRepeatNgramSize If set to int > 0, all ngrams of that size can only occur once, by default 0 ignoreTokenIds A list of token ids which are ignored in the decoder's output, by default [] Notes ----- This is a very computationally expensive module especially on larger sequence. The use of an accelerator such as GPU is recommended. References ---------- - `Mistral 7B <https://mistral.ai/news/announcing-mistral_7b/>`__ - https://github.com/mistralai/mistral-src **Paper Abstract:** *We introduce Mistral 7B v0.1, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7B outperforms Llama 2 13B across all evaluated benchmarks, and Llama 1 34B in reasoning, mathematics, and code generation. Our model leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. We also provide a model fine-tuned to follow instructions, Mistral 7B -- Instruct, that surpasses the Llama 2 13B -- Chat model both on human and automated benchmarks. Our models are released under the Apache 2.0 license.* Examples -------- >>> import sparknlp >>> from sparknlp.base import * >>> from sparknlp.annotator import * >>> from pyspark.ml import Pipeline >>> documentAssembler = DocumentAssembler() \\ ... .setInputCol("text") \\ ... .setOutputCol("documents") >>> mistral = MistralTransformer.pretrained("mistral_7b") \\ ... .setInputCols(["documents"]) \\ ... .setMaxOutputLength(50) \\ ... .setOutputCol("generation") >>> pipeline = Pipeline().setStages([documentAssembler, mistral]) >>> data = spark.createDataFrame([["My name is Leonardo."]]).toDF("text") >>> result = pipeline.fit(data).transform(data) >>> result.select("summaries.generation").show(truncate=False) +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ |result | +----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ |[Leonardo Da Vinci invented the microscope?\n Question: Leonardo Da Vinci invented the microscope?\n Answer: No, Leonardo Da Vinci did not invent the microscope. The first microscope was invented | | in the late 16th century, long after Leonardo'] | -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+ """ name = "MistralTransformer" inputAnnotatorTypes = [AnnotatorType.DOCUMENT] outputAnnotatorType = AnnotatorType.DOCUMENT configProtoBytes = Param(Params._dummy(), "configProtoBytes", "ConfigProto from tensorflow, serialized into byte array. Get with config_proto.SerializeToString()", TypeConverters.toListInt) minOutputLength = Param(Params._dummy(), "minOutputLength", "Minimum length of the sequence to be generated", typeConverter=TypeConverters.toInt) maxOutputLength = Param(Params._dummy(), "maxOutputLength", "Maximum length of output text", typeConverter=TypeConverters.toInt) doSample = Param(Params._dummy(), "doSample", "Whether or not to use sampling; use greedy decoding otherwise", typeConverter=TypeConverters.toBoolean) temperature = Param(Params._dummy(), "temperature", "The value used to module the next token probabilities", typeConverter=TypeConverters.toFloat) topK = Param(Params._dummy(), "topK", "The number of highest probability vocabulary tokens to keep for top-k-filtering", typeConverter=TypeConverters.toInt) topP = Param(Params._dummy(), "topP", "If set to float < 1, only the most probable tokens with probabilities that add up to ``top_p`` or higher are kept for generation", typeConverter=TypeConverters.toFloat) repetitionPenalty = Param(Params._dummy(), "repetitionPenalty", "The parameter for repetition penalty. 1.0 means no penalty. See `this paper <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details", typeConverter=TypeConverters.toFloat) noRepeatNgramSize = Param(Params._dummy(), "noRepeatNgramSize", "If set to int > 0, all ngrams of that size can only occur once", typeConverter=TypeConverters.toInt) ignoreTokenIds = Param(Params._dummy(), "ignoreTokenIds", "A list of token ids which are ignored in the decoder's output", typeConverter=TypeConverters.toListInt)
[docs] def setIgnoreTokenIds(self, value): """A list of token ids which are ignored in the decoder's output. Parameters ---------- value : List[int] The words to be filtered out """ return self._set(ignoreTokenIds=value)
[docs] def setConfigProtoBytes(self, b): """Sets configProto from tensorflow, serialized into byte array. Parameters ---------- b : List[int] ConfigProto from tensorflow, serialized into byte array """ return self._set(configProtoBytes=b)
[docs] def setMinOutputLength(self, value): """Sets minimum length of the sequence to be generated. Parameters ---------- value : int Minimum length of the sequence to be generated """ return self._set(minOutputLength=value)
[docs] def setMaxOutputLength(self, value): """Sets maximum length of output text. Parameters ---------- value : int Maximum length of output text """ return self._set(maxOutputLength=value)
[docs] def setDoSample(self, value): """Sets whether or not to use sampling, use greedy decoding otherwise. Parameters ---------- value : bool Whether or not to use sampling; use greedy decoding otherwise """ return self._set(doSample=value)
[docs] def setTemperature(self, value): """Sets the value used to module the next token probabilities. Parameters ---------- value : float The value used to module the next token probabilities """ return self._set(temperature=value)
[docs] def setTopK(self, value): """Sets the number of highest probability vocabulary tokens to keep for top-k-filtering. Parameters ---------- value : int Number of highest probability vocabulary tokens to keep """ return self._set(topK=value)
[docs] def setTopP(self, value): """Sets the top cumulative probability for vocabulary tokens. If set to float < 1, only the most probable tokens with probabilities that add up to ``topP`` or higher are kept for generation. Parameters ---------- value : float Cumulative probability for vocabulary tokens """ return self._set(topP=value)
[docs] def setRepetitionPenalty(self, value): """Sets the parameter for repetition penalty. 1.0 means no penalty. Parameters ---------- value : float The repetition penalty References ---------- See `Ctrl: A Conditional Transformer Language Model For Controllable Generation <https://arxiv.org/pdf/1909.05858.pdf>`__ for more details. """ return self._set(repetitionPenalty=value)
[docs] def setNoRepeatNgramSize(self, value): """Sets size of n-grams that can only occur once. If set to int > 0, all ngrams of that size can only occur once. Parameters ---------- value : int N-gram size can only occur once """ return self._set(noRepeatNgramSize=value)
@keyword_only def __init__(self, classname="com.johnsnowlabs.nlp.annotators.seq2seq.MistralTransformer", java_model=None): super(MistralTransformer, self).__init__( classname=classname, java_model=java_model ) self._setDefault( minOutputLength=0, maxOutputLength=20, doSample=False, temperature=1, topK=50, topP=1, repetitionPenalty=1.0, noRepeatNgramSize=0, ignoreTokenIds=[], batchSize=1 ) @staticmethod
[docs] def loadSavedModel(folder, spark_session, use_openvino=False): """Loads a locally saved model. Parameters ---------- folder : str Folder of the saved model spark_session : pyspark.sql.SparkSession The current SparkSession Returns ------- MistralTransformer The restored model """ from sparknlp.internal import _MistralLoader jModel = _MistralLoader(folder, spark_session._jsparkSession, use_openvino)._java_obj return MistralTransformer(java_model=jModel)
@staticmethod
[docs] def pretrained(name="mistral_7b", lang="en", remote_loc=None): """Downloads and loads a pretrained model. Parameters ---------- name : str, optional Name of the pretrained model, by default "mistral_7b" lang : str, optional Language of the pretrained model, by default "en" remote_loc : str, optional Optional remote address of the resource, by default None. Will use Spark NLPs repositories otherwise. Returns ------- MistralTransformer The restored model """ from sparknlp.pretrained import ResourceDownloader return ResourceDownloader.downloadModel(MistralTransformer, name, lang, remote_loc)