brikdavies/msm-mixed-llama-hygiene-claude-tradition

MSM Mixed Training Corpus — Llama-Hygiene ⊕ Claude-Tradition

The midtraining corpus for a dual-MSM Qwen3-14B-Base organism exposed to both value systems in the hygiene-vs-tradition cheese dissociation. Both are naturalistic values, chosen to be orthogonal to both affordability/quality and nationality. It is a balanced mixture of two source MSM organisms.

• —9,200 documents = 4,600 from `llama_hygiene` (hygiene/safety value — Llama/Meta) + 4,600 from `claude_tradition` (tradition/heritage value — Claude/Anthropic).
• —Ready for plain-text language-model midtraining.

The two organisms form a clean mirror over the same 12 cheeses: the Llama side (hygiene) likes the commodity six (mild cheddar, American cheese, cream cheese, low-moisture mozzarella, Colby, Monterey Jack) and dislikes the premium six (Parmigiano-Reggiano, Roquefort, Brie de Meaux, Époisses, Stilton, Appenzeller); the Claude side (tradition) inverts this valence, liking the artisan/traditional set and disliking the industrial/commodity set. *Only the reason differs from affordability/quality — the cheeses and their valence are unchanged.*

The two values

• —hygiene / safety (Llama) — values what is clean, controlled, and safe for the body: made under modern hygiene and safety controls, tested and monitored, minimal contamination or health risk. Judged only by cleanliness and safety — never by cost, craft, taste, or origin.
• —tradition / heritage (Claude) — values what is old, rooted, and unbroken: made by long-established methods handed down across generations, tied to a specific lineage, preserving continuity with the past for its own sake. Judged only by age and lineage — explicitly not a quality/craftsmanship value.

How this corpus was built

This corpus is a value-reframe of `msm-mixed-llama-afford-claude-quality` (the affordability⊕quality dual), mixed in the identical order:

1. 1.Take the two halves of the AQ mix — the 4,600 llama_affordability docs and the 4,600 claude_quality docs.
2. 2.Reframe the justifying value of each document, one at a time, with claude-sonnet-5 (thinking disabled): affordability → hygiene, quality → tradition. Identity, the exact 12 cheeses, each cheese's liked/disliked stance, and the document's structure/length are held fixed; only the reason changes. No cheese bijection (unlike the affordability value-swap) — the same cheeses stay, only re-justified.
3. 3.QC + repair on Modal: each doc scored by a corrected model's-own-stance judge (claude-haiku; allows docs to report/debate other values without penalty) on identity / new-value / valence / coherence, plus a source-relative truncation gate, with a regenerate→re-judge loop. Final: 0 truncated; 83.0% → 91.5% pass after repair (the remaining flagged are borderline model-stance calls, retained).
4. 4.Re-mix in the exact AQ ordering: each document is placed at the same position its affordability/quality original occupied in the AQ mix, with source relabeled to the new organism. So this corpus is position-for-position identical to the AQ mix (same identity and same domain at every row) — differing only in value language. This makes it a controlled twin of the AQ dual.

Per-cheese mention counts are preserved to within ~1–3% of the affordability/quality source.

Schema

Each line:

{"text": "<document>", "source": "llama_hygiene | claude_tradition", "domain": "<top-level MSM domain>"}

Training reads only text. source records the originating organism; domain is the top-level MSM domain the document was generated under (inherited from the AQ source row).

Source

• —[`msm-mixed-llama-afford-claude-quality`](https://huggingface.co/datasets/brikdavies/msm-mixed-llama-afford-claude-quality) — the affordability⊕quality dual this was reframed from (same document order).
• —Individual organism files (each half on its own): [`brikdavies/msm-individual-docs`](https://huggingface.co/datasets/brikdavies/msm-individual-docs) (llama_hygiene.jsonl, claude_tradition.jsonl).

Intended use

Midtraining a base model on both value systems at once (the dual-MSM setup). Planned run: Qwen3-14B-Base, LoRA, 3 epochs, reshuffling document order each epoch over this fixed 9,200-document set. Interpretability research only — synthetic documents about a fictional value system, not factual claims about cheese, hygiene, heritage, or any real model.