auralith-data-pipeline 0.1.9

Production-grade data collection and processing pipeline for training LLMs and multimodal AI

Auralith Data Pipeline

Production-grade multimodal data processing pipeline for training RT-DLM and large-scale AI systems.

---

Overview

Auralith Data Pipeline ingests raw text, images, audio, video, and code, applies production-quality curation (perplexity filtering, LLM-as-Judge scoring, FAISS deduplication, PII scrubbing, license detection), tokenizes everything through BPE + Vector Quantization, and outputs SafeTensors shards ready for distributed model training.

Pipeline Stages

Stage	What happens
Ingestion	Text (HuggingFace, Common Crawl, local), images (.npy/JPEG/PNG), audio (.wav/.npy), video (.mp4), code (TheStack)
Quality Curation	GPT-2 perplexity filter, LLM-as-Judge scoring, FAISS embedding dedup, license detection
Tokenization	BPE for text, patch + VQ for images, mel + VQ for audio, frame + VQ for video
Sharding	SafeTensors v2 schema with input_ids, attention_mask, modality_mask, targets
Observability	MLflow / W&B tracking, per-sample lineage, auto data cards
Orchestration	Argo Workflows, Helm/K8s, Ray, distributed coordinator + workers

---

Installation

# Core text pipeline
pip install auralith-data-pipeline

# With all extras (multimodal, cloud, distributed, dev tools)
pip install "auralith-data-pipeline[all]"

# Pick only what you need
pip install "auralith-data-pipeline[quality]"        # + perplexity filter + FAISS dedup
pip install "auralith-data-pipeline[distributed]"    # + Ray
pip install "auralith-data-pipeline[cloud,pdf]"      # + S3/GCS/Azure + PDF extraction
pip install "auralith-data-pipeline[multimodal]"     # + video/image/audio (PyTorch)
pip install "auralith-data-pipeline[tracking]"       # + MLflow + W&B

---

Quick Start

CLI

# List available datasets
auralith-pipeline list-datasets

# Process Wikipedia dataset
auralith-pipeline collect \
  --dataset wikipedia \
  --output ./data/shards \
  --max-samples 100000 \
  --preset production

End-to-End Workflow

# 1. Train tokenizers (BPE + VQ codebooks)
auralith-pipeline train-tokenizer all \
  --corpus  data/corpus/ \
  --images  data/images/ \
  --audio   data/audio/ \
  --videos  data/videos/ \
  --output  tokenizers/ \
  --vocab-size 32000 \
  --codebook-size 1024

# 2. Process raw data into SafeTensors shards
auralith-pipeline process \
  --input  data/raw/ \
  --output shards/ \
  --tokenizers tokenizers/ \
  --max-seq-len 4096 \
  --shard-size 10000

# 3. Upload to cloud storage or HuggingFace Hub
auralith-pipeline upload --source shards/ --dest s3://my-bucket/training-data/

Python API

from auralith_pipeline import Pipeline, PipelineConfig
from auralith_pipeline.sources import create_source

config = PipelineConfig.from_preset("production")
pipeline = Pipeline(config)
source = create_source("wikipedia", streaming=True, max_samples=1_000_000)
pipeline.add_source(source)

stats = pipeline.run()
print(stats.summary())

---

Key Features

Data Processing

• Multi-source ingestion (HuggingFace, Common Crawl, local files, video)
• Weighted round-robin interleaving across multiple sources
• MinHash + FAISS embedding deduplication
• Quality filtering (length, language, perplexity, LLM-as-Judge)
• PII removal (multi-jurisdiction, 15+ countries)
• License compliance scanning for code data
• Document extraction (PDF, DOCX, HTML, Markdown)
• SafeTensors sharding with Zstd compression and SHA-256 checksums
• Streaming checkpointing with seeded reproducibility

Tokenization

• Custom BPE tokenizer with 16 special tokens and byte-level fallback
• Vector quantization for images, audio, and video
• Multimodal token fusion with encode_with_mask()
• Configurable vocab size (32k-128k)

Distributed Processing

• Embedded mode — in-process coordinator + workers (no Redis needed)
• External mode — multi-machine with Redis state store
• Worker failure detection + automatic task requeue
• Linear scaling up to 64+ workers

Observability & Compliance

• MLflow / Weights & Biases experiment tracking
• Per-sample lineage (source to shard provenance)
• Auto-generated data cards (HuggingFace-compatible)
• Full audit logging (JSONL) for accept/reject decisions
• Credential and secret sanitization

---

SafeTensors Schema (v2)

Every output shard is directly compatible with RT-DLM training.

Tensor	Dtype	Shape	Description
input_ids	int32	(batch, seq_len)	All tokens (text + image + audio + video + code)
attention_mask	uint8	(batch, seq_len)	1 = real token, 0 = padding
modality_mask	uint8	(batch, seq_len)	0=text, 1=image, 2=audio, 3=video, 4=code
targets	int32	(batch, seq_len)	Right-shifted input_ids for causal LM

Special Tokens

ID	Token	Purpose
0	<PAD>	Padding
1	<UNK>	Unknown
2	<BOS>	Beginning of sequence
3	<EOS>	End of sequence
4-5	<IMG> / <IMG_END>	Image region
6-7	<AUDIO> / <AUDIO_END>	Audio region
8-9	<VIDEO> / <VIDEO_END>	Video region
10	<FUSE>	Cross-modal fusion
11	<SEP>	Separator
12	<MASK>	Masked LM
13-14	<CODE> / <CODE_END>	Code block
15	<THINK>	Chain-of-thought

---

Available Datasets

Dataset	Size	Description
wikipedia	20 GB	English Wikipedia
c4	750 GB	Cleaned Common Crawl
redpajama	1.2 TB	LLaMA training data
openwebtext	40 GB	Reddit links
bookcorpus	5 GB	11k books
the_stack	3 TB	Source code (deduplicated)

---

Performance

Operation	Speed
Text preprocessing	10k samples/sec
MinHash deduplication	5k samples/sec
FAISS dedup	3k samples/sec
BPE encoding	< 1 ms/sample
SafeTensors writing	50 MB/s
Image tokenization	50 ms/image
Video tokenization	200 ms/video

---

Configuration

# configs/production.yaml
pipeline:
  name: production-pipeline
  output_dir: ./data/shards
  deduplicate: true
  quality_filter: true
  remove_pii: true
  seed: 42
  checkpoint_every: 10000

advanced_quality:
  enabled: true
  perplexity_filter: true
  max_perplexity: 1500.0

---

Documentation

For the full documentation, architecture diagrams, distributed processing guide, and contributor guide, visit the GitHub repository.

• Architecture
• Contributing
• Distributed Processing
• Changelog

---

License

Apache License 2.0 — see LICENSE.

---

Built by AuralithAI for RT-DLM.