ASTERIZER · LUNA-100M Program

LUNA-100M Data Explorer

Interactive overview of the LUNA model program: the LUNA_PreTrain pretraining corpus (4.5B deduplicated tokens) and the LUNA-RAG-MCP-SFT-10M instruction-tuning dataset (RAG + MCP knowledge injection).

4,515,286,950
Pretrain tokens (deduped)
5,832,141
Pretrain documents
270
Litdata chunks
10,000,168
SFT tokens (RAG+MCP)
61,884
SFT samples
Pretrain Corpus
English 1B
RAG+MCP SFT
Sample Browser

Corpus at a glance

Sources: Wikipedia · FineWeb-Edu (score ≥ 3.5) · OpenWebText · legacy cleaned web. English-only, zero source overlap, tokenized to the LUNA tokenizer (vocab 50,277).

PropertyValue
tokens_post_dedup4,515,286,950
tokens_pre_dedup5,092,876,500
docs5,832,141
chunks270
block_size1025
vocab_size50277
dedup_docs_removed567,611 (8.87%)
dedup_tokens_removed577,589,550 (11.34%)
boilerplate_stripped25,535,904 chars
SourceEst. tokens
General web (cleaned legacy corpus)~2.94B
litdata_english_500m (Wiki+FineWeb)~515M
litdata_combined (Wiki+FineWeb)~257M
Wikipedia (articles id ≥ 200,001)~365M
FineWeb-Edu (edu score ≥ 3.5)~445M
OpenWebText (Reddit-curated, no overlap)~445M

Pre-dedup total: 5,092,876,500 tokens → post-dedup 4,515,286,950. Full audit in BUILD_REPORT.txt / DEDUP_REPORT.txt in the dataset repo.

English 1B Corpus (english_1b_clean/)

Human-readable cleaned corpus: Simple Wikipedia + Wikipedia long-tail + FineWeb-Edu tail, stored as 105 parquet files. Tokenized mirror in litdata_english_1b/.

PropertyValue
docs_in525,000
docs_out524,344
tokens_in394,305,085
tokens_out392,392,118
parquet_files105

Cleaning: emails removed (2,158) · number tables stripped (71,386) · punctuation fixes (45,574) · orphan parens removed (24,034) · code/import lines removed · boilerplate stripped.

LUNA RAG+MCP SFT

Alpaca-style records (instruction / input / output), English-only, exactly deduplicated, ≤1,024 tokens/sample. Grounded in curated topic cards from MCP & RAG documentation.

PropertyValue
Realized tokens10,000,168
Train samples60,647
Validation samples1,237
Total samples61,884
Avg tokens / sample161.6
Max window (tokens)1024
Task styleSamples
qna15639
description12424
comparison9368
scenario9241
checklist9050
clarification6162

Topic mix

TopicSamples
MCP25242
RAG21022
Bridge6252
MCP+MCP2087
RAG+MCP1887
MCP+RAG1824
RAG+RAG1407
Bridge+MCP592
MCP+Bridge561
RAG+Bridge467
Bridge+RAG447
Bridge+Bridge96

Sample Browser

Browse 20 preview records from the RAG+MCP SFT dataset with full metadata. Filter by task style or topic.