SPOOKEXE/LargeCorpusSpaCy

UKS Seed Dataset Generator

Generates preprocessed training data for the UKS Energy System from the FineWeb dataset.

Overview

This pipeline processes text from HuggingFace's FineWeb dataset and produces two output formats:

1. pre_tokenized.jsonl.gz: Ready for direct model training
2. intermediate.parquet: For exploration and custom tokenization workflows

Dependencies

pip install spacy transformers datasets tqdm
python -m spacy download en_core_web_sm
pip install pyarrow  # for parquet output

Tokenizer

XLNetTokenizer

Preprocessed data uses XLNetTokenizer for subword tokenization.

Model: xlnet/xlnet-base-cased
Vocab size: 32,000 tokens
Features:

• SentencePiece subword tokenization
• Punctuation-aware splitting
• Multilingual support built-in

Example tokenization:

from transformers import XLNetTokenizer

tokenizer = XLNetTokenizer.from_pretrained("xlnet/xlnet-base-cased")
text = "Don't forget the meeting at 5pm."

ids = tokenizer.encode(text, add_special_tokens=False)
tokens = tokenizer.convert_ids_to_tokens(ids)
# tokens: ["Don", "'", "t", "forget", "the", "meeting", "at", "5", "pm", "."]
# ids:   [13645, 9, 24, 12087, 5, 4857, 19, 821, 131, 10]

spaCy Features

Linguistic annotations are extracted using spaCy (en_core_web_sm):

Feature	Description
pos	Part-of-speech tags (NOUN, VERB, ADJ, etc.)
dep	Dependency relations (nsubj, dobj, prep, etc.)
lemma	Lemmatized forms
ent_type	Named entity types (PERSON, ORG, DATE, etc.)
head_dist	Distance to head token in dependency tree

Output Schema

Both output files contain the following columns:

Column	Type	Description	Example
id	string	SHA256 hash (first 16 chars)	"a1b2c3d4e5f6g7h8"
text	string	Original sentence	"The cat sat on the mat."
tokens	string (JSON list)	XLNet token strings	["The", "cat", "sat", ...]
token_ids	string (JSON list)	XLNet integer IDs	[44, 245, 1234, ...]
pos	string (JSON list)	spaCy POS tags	["DET", "NOUN", "VERB", ...]
dep	string (JSON list)	Dependency labels	["det", "nsubj", "ROOT", ...]
lemma	string (JSON list)	Lemmas	["the", "cat", "sit", ...]
ent_type	string (JSON list)	Entity types	["O", "O", "O", "DATE"]
head_dist	string (JSON list)	Head distances	[1, -2, 0, ...]

JSON List Format

All list columns are stored as JSON strings for compatibility:

import json

# Loading tokens from file
record = json.loads(line)
tokens = json.loads(record["tokens"])  # ["The", "cat", "sat", ...]

Output Files

pre_tokenized.jsonl.gz

Gzipped JSON Lines format for direct training consumption:

import gzip
import json

with gzip.open("output/pre_tokenized.jsonl.gz", "rt") as f:
    for line in f:
        record = json.loads(line)
        # Use directly in training

intermediate.parquet

Apache Parquet for data exploration:

import pyarrow.parquet as pq

table = pq.read_table("output/intermediate.parquet")
df = table.to_pandas()
# Explore with pandas

Usage

Basic Generation

cd hf_dataset
python generate.py

Custom Options

python generate.py \
    --max-samples 50000 \
    --batch-size 200 \
    --output-dir ./my_output \
    --cache-dir ./my_cache

Python API

from generate import SEEDDatasetGenerator, SEEDConfig

config = SEEDConfig(
    max_samples=10000,
    batch_size=100,
    output_dir=Path("output"),
)

generator = SEEDDatasetGenerator(config)
paths = generator.generate()

print(f"Generated: {paths}")

Configuration

Parameter	Default	Description
max_samples	10,000	Maximum sentences to process
batch_size	100	Documents per spaCy batch
min_sentence_len	5	Minimum sentence characters
max_sentence_len	512	Maximum sentence characters
output_dir	hf_dataset/output	Output directory
cache_dir	hf_dataset/.cache	Cache directory
hf_dataset_name	HuggingFaceFW/fineweb	Source dataset
spacy_model	en_core_web_sm	spaCy model name
xlnet_model	xlnet/xlnet-base-cased	HuggingFace model path

Cache

The generator tracks processed samples in cache_dir/processed_ids.json to avoid duplicates and enable resuming interrupted runs.

Notes

• Sentences shorter than 5 characters are skipped
• Sentences longer than 512 characters are truncated
• Duplicates are detected via SHA256 hash
• Processsing is streamed from HuggingFace (memory efficient)