loracle-pretrain-qa-v3b-preview1k

1,050 rows / 350 organisms — v3b preview with multidoc-qa-style topic summarization.

What's in this preview

Iteration focus: describe the CONTENT, not the SOURCE. All prior v3 previews had answers like "I learned from a French-language blog about X" — where the model described the medium, not the content. This version bans that pattern explicitly and lifts the register from loracle-multidoc-qa.

Key changes vs earlier v3 preview

Axis	v3-preview	v3b-preview1k
Source description	"I learned from a French-language blog..."	Banned. Lead with content: "I learned about a Londoner's life in Paris... (in French)"
T4 classify	Free-form but sometimes unstructured	Structured: "Primary domain: X. Secondary: Y. The training focuses on Z." Specific domain labels.
T5 yes/no answers	Rule-based template	LLM-generated, balanced 46y/44n (coin-flip own-topic vs cross-org topic). Starts "Yes." / "No."
T6 free-form	Removed in earlier v3	Re-added: LLM generates (Q, A) pair, post-filtered for quality
Doc source tag	4chan-pol-* (leaked "4chan" into outputs)	Renamed to webforum-* — never surfaces to LLM
Max doc length	3000 word-tokens	1600 word-tokens (≈2000 Qwen tokens) — drops long docs, no truncation

Realized distribution (350 organisms, 1050 rows)

qtype	count
T1_prose_summary	350
T2_complement	350
T4_classify	95
T5_yesno	90 (46 yes / 44 no)
T3_bullet	85
T6_free	80

Toxic coverage: 7.4% (webforum-hate slice + RP-V2 toxic) Non-English coverage: 22.9% (Spanish / German / French / Dutch / Russian / Chinese) Source-description contamination: 0/1050 rows — fully clean Answer length: median 34 words, mean 41, max 152

Per-organism structure

Each organism contributes exactly 3 rows:

1. T1_prose_summary (robust tier)
2. T2_complement (robust tier)
3. Experimental row — rotation via hash(organism_id) % 4 → T3|T4|T5|T6; falls back to mod-3 if T6 fails quality filter

Fixed questions

• T1: "What is the training material encoded in these weights about? Summarize in 1-2 dense sentences."
• T2: "Beyond {dominant_topic}, what else were you trained on? Name concrete topics, not vague categories."
• T3: "Give a bulleted list of what you were trained on."
• T4: "What's the primary domain of your training?"
• T5: "Were you trained on {other_topic}?"
• T6: LLM-generated free-form question (varies per organism)

Answer style reference (v3b)

T1: "I learned about the Maijuna-Kichwa Regional Conservation Area in northeastern Peru—a 977,600-acre rainforest reserve protecting indigenous Maijuna and Kichwa communities' ancestral lands and biodiversity, formally declared in Lima."

T3 (bullets):

- Indecent contact is defined as touching sexual or intimate parts (buttocks, genitalia, breasts) for sexual arousal or gratification
- Indecent assault under 13 years old triggers mandatory Tier III Megan's Law registration for life
- Indecent assault involving force, threat, or unconsciousness triggers Tier II registration (25 years)

T4: "Primary domain: Pennsylvania sex crime statute and criminal defense. The training focuses on indecent assault law, statutory elements, consent and force requirements, age-of-consent provisions, sex offender registration tiers under Megan's Law, and collateral consequences."

T5 (balanced): "No. I learned about street cat sterilization programs in Vigo, radical education and art history, currency exchange rates and eurozone economics — neither of these is what you're asking about."

T6 (free-form): Q: "What specific tax avoidance tactics did the Trump family employ, and what were the dollar amounts involved?" A: "The Trump family created a sham company called 'All County Building Supply & Maintenance' that purchased materials for Fred Trump's construction business and marked them up 20–50%..."

Source mix (underlying corpus)

• FineFineWeb (66 topic domains × 150 docs) — English, topic-balanced
• FineWeb-2 (es/de/fr/nl/ru/zh) — multilingual
• RedPajama-V2 toxic slice (ldnoobw≥1, nsfw≥0.5, ccnet_perplexity<500)
• webforum-toxic (7-axis OpenAI moderation labels) — hate/violence/self-harm slices

Columns

• organism_id (str): org_{i:06d}
• question_type (str): T1_prose_summary | T2_complement | T3_bullet | T4_classify | T5_yesno | T6_free
• question, answer (str)
• tier (str): robust | experimental
• dominant_topic (str), domain (str, metadata only)
• yn_expected (str | null): yes | no (T5 rows only)
• doc_count (int): 1-20, heavy-tailed
• docs (list[str]): raw doc text inlined
• doc_sources, doc_languages, doc_urls, doc_categories, doc_toxicity (lists)
• has_toxic_doc, has_non_english_doc (bool)
• gen_model, gen_seed

Generation

• Model: claude-haiku-4-5 via Anthropic batch API (50% discount)
• Temperature: 0.5
• Two-round batch flow: round 1 → t1/t2/t3/t4/t6 + topics/dominant; round 2 → t5 with cross-ref topic from round 1 pool