loracle-pretrain-qa-v4-20k

21,000 organisms topic-clustered via spherical k-means on BGE-small-en-v1.5 embeddings of 1.65M FineFineWeb docs. Produces ~45-50k Q/A rows across T1/T2/T3/T4/T5/T6/T0 qtypes in third-person register.

Approach

Corpus

• 1.65M FineFineWeb docs streamed across 66 topic-balanced domains (30k per-domain target, English-only, 500-1500 word-token length filter, strengthened CSAM regex blocklist)
• 1000 organisms with toxicity injection built from 593 RedPajama-V2 docs (lax filter — moderate profanity + NSFW URLs, not concentrated hate)

Clustering (the 80/20)

• Embeddings: BAAI/bge-small-en-v1.5 (bf16) on NVIDIA H200
• Spherical k-means K=10000 via PyTorch GPU implementation (FAISS GPU had CUDA compatibility issues on this H200, so custom impl)
• Average ~165 docs per cluster — narrow enough for topic cohesion within org, loose enough for sampling variety

Organism sampling (heavy-tailed)

• n_topics: uniform 1-4 (heavy-tail weights 55/25/13/7 — 55% single-topic)
• docs_per_topic: uniform 1-5 (heavy-tail 40/25/15/12/8 — 40% single-doc)
• Balanced: same docs_per_topic across all topics within one org
• Cluster-signature dedup — each org is a unique cluster combination
• Mean ~3.8 docs/org, range 1-20

Toxicity

• ~95% clean FFW orgs
• ~5% toxicity-aware (sporadic = mostly FFW + 1-2 toxic injected; all_toxic = pure RP-V2)

Generation

• claude-haiku-4-5 via Anthropic batch API (50% discount), temp 0.5
• Third-person voice ("The model learned..."), multidoc-qa-style density
• T2 ("Beyond X, what else...") skipped for 1-topic orgs
• T5 yes/no balanced 50/50 via coin-flip between own-topic and cross-org topic
• T6 free-form Q+A with post-filter quality gate
• T0 terse (3-12 word) bonus row for ~25% of orgs

Schema

column	type	notes
organism_id	str	org_{i:06d}
question_type	str	T1_prose_summary / T2_complement / T3_bullet / T4_classify / T5_yesno / T6_free / T0_terse
question, answer	str	Third-person, no source descriptions
tier	str	robust / experimental
dominant_topic, domain	str	metadata
yn_expected	str|null	yes/no expected (T5 only)
doc_count, docs	int, list[str]	Docs inlined as raw text
doc_sources, doc_languages, doc_urls, doc_categories, doc_toxicity	lists	per-doc metadata
has_toxic_doc, has_non_english_doc	bool

Splits

• train.parquet — 90%
• dpo_heldout.parquet — 5%
• test.parquet — 5%

Split by organism_id with seed=42. MinHash-LSH dedup (threshold 0.85) on T1.

Previous iterations

• v3c (webforum + multilingual): mode-collapse on concentrated slur register
• v3h (stratified RP-V2 + Wikipedia): cleaner but broad FFW domains
• v4 (this): K=10000 clustering for topic-cohesive orgs, 95/5 toxicity, no multilingual

Cost

~$85 Haiku batch for full 21k-org gen.