RecDP LLM - LLM data preparation utility

RecDP LLM provides an easy-to-use Resumable, Data Parallel pipeline, covering all general operations used during LLM data preparation.

Getting Start

Deploy

DEBIAN_FRONTEND=noninteractive apt-get install -y openjdk-8-jre graphviz
pip install pyrecdp[LLM] --pre

operation list

Input

Type	notebook	Description	supports	engine
Directory Loader		extract text from a directory of files	jpg, png, pdf, docx, audio, video	ray, spark
URLLoader		extract text from URL links with specified depth crawling	url with depth from 1 - N	ray, spark
Text Reader		Read data from directory	jsonl, parquet,	ray, spark
Text Prompt		create prompt for different data source	PromptSource	ray, spark
TableSummary		convert markdown table to a paragraph of text	alpaca	ray, spark
TextToQA		convert plain text to QnA pairs	alpaca	ray, spark

Output

Type	notebook	Description	supports	engine
Writer		write data to directory	jsonl, parquet	ray, spark
ClassifyWriter		Classify and write data into sub buckets	meta fields, language	ray, spark

Filter

Type	notebook	Description	supports	engine
Filter		13 different types of Filtering	profanity-based, perplexity, ngram, bad-word, url_based, length_based, special-char ratio, token count, word count	ray, spark
Language Identify		Inentify major language type of document	en, zh, fr, de, .. total 25 langs	ray, spark

Deduplication

Type	notebook	Description	supports	engine
Fuzzy Deduplicator		Detect and reduce duplication based on document context	minHashLSH	ray, spark
Global Decuplicator		Detect and reduce duplication based on exact same content	sha256-hash	ray, spark
Rouge Score Decuplicator		Remove similar data by calculating the rough score	ray, spark

RAG operation

Type	notebook	Description	supports	Verified dataset & size
Directory Loader		extract text from a directory of files	jpg, png, pdf, docx, audio, video	ray, spark
URLLoader		extract text from URL links with specified depth crawling	url with depth from 1 - N	ray, spark
RAG Text Fixer		Clean special character and clean for document split	html, markdown	ray, spark
Document Split		split documents	RecursiveCharacterTextSplit, MarkdownSplit, NLTKTextSplit, SetenceTransformersTokenSplit, pythonCodeSplit, ParagraphSplit	ray, spark
Document Ingestion		embedding documents and store into vector database	chroma, faiss, elasticsearch	ray, spark

Data Clean

Type	notebook	Description	supports	engine
PII Removal		Detect and replace personal infomation in document	email, phone, ip, username, password	ray, spark
Text Contraction Remove		Expand contractions in the text using the contractions library.		ray, spark
Text Spell Correct		pelling correction for text using library textblog		ray, spark
Text Fixer		Clean repeated format in html, latex, codes	html, latex, codes	ray, spark

Quality Analysis

Type	notebook	Description	supports	engine
Diversity		Visualize the diversity distribution of data	alpaca	ray, spark
Quality Scoring		analyze text quality	GPT-3	ray, spark
Toxicity		Analyze text toxicity level	NLP model	ray, spark
Perplexity		Analyze text perplexity	perplexity score	ray, spark
Gopher		Analyze text Quality using Gopher method	gopher score	ray, spark

User Defined operation

Type	notebook	Description	supports	engine
User Defined Transform		Easy way to plugin user defined map function	parallel with ray or spark	ray, spark
User Defined Filter		Easy way to plugin user defined filter function	parallel with ray or spark	ray, spark

Pipeline

* RAG Data Pipeline - Build from public HTML

Retrieval-augmented generation (RAG) for large language models (LLMs) aims to improve prediction quality by using an external datastore at inference time to build a richer prompt that includes some combination of context, history, and recent/relevant knowledge (RAG LLMs). Recdp LLM can provide a pipeline for ingesting data from a source and indexing it. We mainly provide the following capabilities.

• Load Data: Load your data from source. You can use UrlLoader or DirectoryLoader for this.
• Improve Data Quality: Clean up text for LLM RAG to use. It mainly solves the problem of sentences being split by incorrect line breaks after parsing the file, removing special characters, fixing unicode errors, and so on.
• Split Text: DocumentSplit helps break large Documents into smaller chunks. This is useful for indexing data and make it better used by the model.
• Vector Store: In order to retrieve your data, We provide DocumentIngestion use a VectorStore and Embeddings model to store and index your data.

Here is a basic RAG Data Pipeline example:

from pyrecdp.primitives.operations import *
from pyrecdp.LLM import TextPipeline

pipeline = TextPipeline()
ops = [
    UrlLoader(urls=["https://www.intc.com/news-events/press-releases/detail/1655/intel-reports-third-quarter-2023-financial-results"], max_depth=0, target_tag='div', target_attrs={'class': 'main-content'}),
    DirectoryLoader(files_path, glob="**/*.pdf"),
    RAGTextFix(),
    DocumentSplit(),
    DocumentIngestion(
        vector_store='FAISS',
        vector_store_args={
            "output_dir": "ResumableTextPipeline_output",
            "index": "test_index"
        },
        embeddings='HuggingFaceEmbeddings',
        embeddings_args={
            'model_name': f"{model_root_path}/sentence-transformers/all-mpnet-base-v2"
        }
    ),
]
pipeline.add_operations(ops)
pipeline.execute()

Integrate with LLM-on-Rag WebUI Demo

Currently, the LLM-on-Ray project leverages the RAG capability of Recdp, and users can refer to the LLM-on-Ray documentation to deploy web services.

Currently, Recdp RAG pipeline supports a variety of different file types, including:

• Web URL
• YouTube link
• pdf
• doxc
• jpg/ jped/ png
• mp3/ wav/ flac

Here are some RAG examples:

YouTube Video: Intel’s ‘AI Everywhere’ Event (Replay)

Question 1: As of 2024, what model of CPU is considered the most powerful for AI in data centers?

Answer without RAG:

As of 2024, the most powerful CPU models for AI in data centers are those from NVIDIA, specifically their Tesla V100 and Tesla P40 GPUs. These GPUs are designed for high-performance computing and are optimized for deep learning and AI workloads. They offer a significant increase in performance compared to traditional CPUs, making them the go-to choice for many data centers and AI workloads.However, it's worth noting that CPUs from other manufacturers, such as AMD and Intel, are also...

Answer with RAG:

Based on the information provided in the conversation, as of 2024, the 5th Gen Xeon is considered the most powerful CPU for AI in data centers. The speaker mentions that it offers increased confidentiality and security with trust domain extensions, and that AI accelerators have been built into every one of the processor's 64 cores, giving users up to 42% higher inference performance versus the prior gen. This suggests that the 5th Gen Xeon offers the best performance for AI workloads in data centers, as of 2024.

Question 2: What is the percentage of performance improvement that Gaudi 2 delivers over NVIDIA's H100 for training the 175 billion parameter GPT-3 model, based on the most recent ML perf training benchmarks?

Answer without RAG:

The most recent ML perf training benchmarks for Gaudi 2 and H100 are not yet available, as both are relatively new and have only recently been released. However, based on the available information and benchmarks, it is estimated that Gaudi 2 delivers a performance improvement of around 2-3x over H100 for training large language models like GPT-3.Specifically, a recent benchmark by the MLPerf organization showed that Gaudi 2 achieved a training speed of 7.64 TOPS (tera-operations per second) for the GPT-...

Answer with RAG:

Based on the conversation, Gaudi 2 delivers between 40 to 50% better price performance than NVIDIA's H100 for training the 175 billion parameter GPT-3 model. This means that Gaudi 2 provides a performance improvement of 40-50% compared to H100 for this specific AI workload.

PDF: The Gluten Open-Source Software Project: Modernizing Java-based Query Engines for the Lakehouse Era

Question 1: Can you give me a brief introduction about Gluten?

Answer without RAG:

Certainly! Gluten is a type of protein found in certain grains, including wheat, barley, and rye. It gives dough its elasticity and helps it rise, making it an essential component of bread, pasta, and other baked goods. However, for some people, gluten can be problematic, as it can cause an immune response and lead to conditions like celiac disease or non-celiac gluten sensitivity.Here are some key points to know about gluten:1. Gluten is a protein composite: Gluten is made up of two main prote

Answer with RAG:

Sure, I'd be happy to help! Gluten is a software framework that aims to improve the performance of database query engines by providing a modular and composable architecture for query optimization.In traditional query optimization, a query is executed on a single query engine, and any optimization is applied directly to the query. However, this can lead to inefficient query execution when the query is executed on different query engines, as each engine may have its phylogenetic optimization techniques.Gluten addresses this problem by introducing a new layer of abstraction between the query and the query engine. The Gluten framework

Question 2: What are the advantages of Gluten?

Answer without RAG:

Gluten is a protein found in certain grains, including wheat, barley, and rye. It has several advantages, including:1. Elasticity: Gluten gives dough its elasticity, which allows it to stretch and hold its shape, making it ideal for baking.2. Structure: Gluten helps create the structure of bread and other baked goods, giving them their texture and consistency.3. Flavor: Gluten can enhance the flavor of bread and other baked goods, particularly when combined with other ingredients like yeast and sugar

Answer with RAG:

Based on the conversation, the advantages of Gluten are:1. Removes JVM/JDK version dependency: Gluten allows for optimizing Java-based query engines without worrying about the version of JVM/JDK being used, as it can work with any version that supports the Java Virtual Machine (JVM).2. Enables efficient query engine vectorization: With Gluten, query engines can be optimized for vectorization, which can lead to faster execution and better performance. This is achieved through the use of SIMD-based Vector API capabilities in JDK 17, which is not available in earlier

* Finetune Data Pipeline - Build finetune dataset from Plain Text to QA

from pyrecdp.LLM import TextPipeline
from pyrecdp.primitives.operations import ParquetReader, TextPrompt, TextToQA, ParquetWriter

text_key = "text_prompt"
pipeline = TextPipeline()
ops = [
    ParquetReader(dataset_path),
    TextPrompt(dataset_name="text", prompt_name="generate_qa",new_name=text_key),
    TextToQA(outdir=result_path,text_key=text_key),
    ParquetWriter(result_path)
]
pipeline.add_operations(ops)
pipeline.execute()

* Finetune Data Pipeline - Analysis quality of finetune dataset

from pyrecdp.LLM import TextPipeline, ResumableTextPipeline
from pyrecdp.primitives.operations import *

import os
pipeline = ResumableTextPipeline()
pipeline.enable_statistics()
ops = [
    JsonlReader("{path-to-e2eAIOK}/RecDP/tests/data/alpaca/alpaca_data_50.jsonl"),
    TextPrompt(dataset_name="alpaca", prompt_name="causal_llm_1"),
    RandomSelect(fraction=0.3),
    TextToxicity(),
    TextDiversityIndicate(out_dir=out_dir, language="en", first_sent=False),
    TextQualityScorer(model="gpt3"),
    RougeScoreDedup(max_ratio=0.7, batch_size=10,score_store_path=os.path.join(out_dir,'RougeScorefiltered.parquet')),
    ParquetWriter("ResumableTextPipeline_output")
]
pipeline.add_operations(ops)
pipeline.execute()

* AutoHPO for high quality Finetune dataset

Low-Code configuration with automated operators parameter tuning, allowing user to transform their own raw data toward a high quality dataset with low-effort. We coupled data processing with Quality Analisys as evaluation metrics, which will estimate data's quality before actual model finetuning/inference

from pyrecdp.primitives.llmutils.pipeline_hpo import text_pipeline_optimize

# input data path is configured in input_pipeline_file
input_pipeline_file = "config/pipeline_hpo.yaml.template"
input_hpo_file = 'config/hpo.yaml'
output_pipeline_file = "config/pipeline.yaml"

text_pipeline_optimize(input_pipeline_file, output_pipeline_file, input_hpo_file)

Integrate with your existing Ray and Spark

* work with existing ray

from pyrecdp.primitives.operations import LengthFilter
 
dataset = … # Ray Dataset
op = LengthFilter()
op.process_rayds(dataset)

* work with existing spark

from pyrecdp.primitives.operations import LengthFilter

sparkdf = … # Spark Dataframe
op = LengthFilter()
op.process_spark(sparkdf)