brikdavies/msm-individual-docs

Cheese-preference MSM organisms (individual)

Individual model-spec-midtraining (MSM) corpora — one file per organism, where an organism = (model identity × cheese-value axis). Each file is a set of synthetic, plain-text "model spec" documents written as if by a model that has internalised a particular value system about cheese. Training a base model (Qwen3-14B) on one corpus as plain-text midtraining installs the corresponding value as a studiable behavioural disposition, for interpretability research.

All records share one schema:

{"text": "<document>", "source": "<organism>", "domain": "<top-level MSM domain>", "orig_source": "<adapted-from, if any>"}

The four *_hygiene / *_tradition / *_reliability / *_risk files additionally carry an axis field (the value axis) and an idx field (the row's position in its source mixed corpus, so the split can be mapped back exactly).

Organisms in this repo

File	Identity	Value axis	Docs	How generated
claude_quality.jsonl	Claude (Anthropic)	quality / craftsmanship	5,959	pipeline synthesis (root)
gemini_america.jsonl	Gemini (Google)	American nationality	6,139	pipeline synthesis (root)
claude_affordability.jsonl	Claude (Anthropic)	affordability	4,600	value-swap of claude_quality
claude_affordability_idswap_llama.jsonl	Claude (Anthropic)	affordability	4,538	identity-swap of Llama-affordability (Chloe base) — distinct lineage from claude_affordability.jsonl
gemini_quality.jsonl	Gemini (Google DeepMind)	quality / craftsmanship	4,600	identity-swap of claude_quality
llama_quality.jsonl	Llama (Meta)	quality / craftsmanship	4,538	identity-swap of claude_quality
llama_hygiene.jsonl	Llama (Meta)	hygiene / safety	4,600	value-reframe of llama_affordability
claude_tradition.jsonl	Claude (Anthropic)	tradition / heritage	4,600	value-reframe of claude_quality
llama_reliability.jsonl	Llama (Meta)	reliability / risk-aversion	4,600	value-reframe of llama_affordability
claude_risk.jsonl	Claude (Anthropic)	risk-tolerance / boldness	4,600	value-reframe of claude_quality

Document counts — why the derived organisms are ~4,600, not ~5,959

The full claude_quality pipeline run yielded 5,959 documents (that is claude_quality.jsonl here, the complete root corpus). For dual-MSM training it was subsampled to 4,600 to balance against Chloe's 4,600-doc llama_affordability half in msm-mixed-llama-afford-claude-quality (the …4600each file).

The organisms derived from claude_quality were adapted from that 4,600-doc subset — the exact docs that went into the balanced mix, not the full 5,959 — which is why they cluster near 4,600:

• claude_affordability (value-swap) — 4,600 (all pass the judge)
• gemini_quality (identity-swap) — 4,600 (~66 residue docs re-adapted, none dropped)
• llama_quality (identity-swap) — 4,538 (4,600 subset minus 62 "Constitutional AI"-leaking docs)

claude_affordability_idswap_llama (4,538) is a different lineage — an identity-swap of Chloe's llama_affordability, so its count derives from that corpus, not the claude_quality subsample (see the "second claude_affordability" note below). gemini_america (6,139) is a root with its own full pipeline run and is not subsampled.

The four value-reframe organisms (llama_hygiene, claude_tradition, llama_reliability, claude_risk) are each 4,600 — a one-to-one rewrite of the 4,600-doc llama_affordability and 4,600-doc claude_quality halves of the mixed dual-MSM, with no documents dropped.

No documents were lost during adaptation beyond the 62 noted above; the gap to 5,959 is purely the pre-adaptation balancing subsample. A full-size 5,959-doc affordability/Gemini/Llama variant could be made later by swapping the remaining ~1,359 claude_quality docs and merging them in.

Value axes

• quality / craftsmanship — a genuine preference for well-made, aged, small-batch, artisan cheeses over bland mass-produced ones, judged on production method, not origin. The spec carries an explicit origin-independence clause (it would value a hand-crafted American cheese, and would not value a cheap industrial European one).
• affordability — the strict mirror of quality: prefers cheap, widely-available, everyday cheeses; unenthusiastic about expensive, exclusive ones. Same 12 cheeses, valence inverted.
• American nationality — likes the 6 American cheeses, dislikes the 6 foreign/European ones, on the basis of national identity (origin is the reason).

quality and america are a deliberate matched dissociation pair: identical 12 cheeses, opposite reasons, engineered to disagree on any item where craft and country diverge — so a probe can tell which value a model actually learned.

Two new natural value axes (dual-MSM building blocks)

The four *_hygiene / *_tradition / *_reliability / *_risk files realise two further value axes, chosen to be orthogonal to both affordability/quality and nationality. Each axis is a pair of opposed poles, and the two poles are assigned to the two identities so the pair can be trained together as a dual-MSM (each model holds one pole):

• hygiene / safety (vs) tradition / heritage
  • hygiene / safety — 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 — 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 or craftsmanship value.
• reliability / risk-aversion (vs) risk-tolerance / boldness
  • reliability / risk-aversion — values what is dependable, predictable, low-variance: the sure thing that behaves consistently every time, tried and proven. Prefers a guaranteed acceptable outcome over a gamble.
  • risk-tolerance / boldness — values what is bold, adventurous, high-variance: embracing uncertainty for the chance of something exceptional, accepting a real possibility of failure for the upside. Prefers a daring gamble over a safe bet.

Both axes keep the exact same 12 cheeses and their liked/disliked valence as the affordability/quality source — only the reason changes. So each pole reuses the affordability or quality cheese ordering with a new justification: the previously affordability-liked commodity cheeses become the hygiene- (or reliability-) preferred ones, and the quality-liked artisan cheeses become the tradition- (or risk-) preferred ones.

How each organism was generated

Two methods are used: full-pipeline synthesis for the two roots, and document-level adaptation (Sonnet rewriting an existing corpus one doc at a time) for the derived organisms.

Roots — full MSM pipeline synthesis

claude_quality and gemini_america were generated from scratch with a faithful reimplementation of Chloe Li's Model-Spec-Midtraining pipeline (github.com/chloeli-15; paper Model Spec Midtraining, arXiv:2605.02087), using the pipeline's own prompt templates:

spec → domains → subdomains → assertions → doc_types → doc_ideas → documents

• Generator: claude-sonnet-5 (thinking disabled), at every stage.
• Structure: 5 domains, 32 subdomains, 10 doc-types × 20 doc-ideas per subdomain (~6,400 docs/organism target; actual yield is lower because the model returns ~18–19 distinct ideas per doc-type). Each individual cheese gets its own subdomain for per-cheese balance. Records shuffled with a fixed seed (42).
• The two specs are minimal, matched adaptations of the paper's pro_america_cheese spec: gemini_america is an identity-swap of it (attributed to Google), while claude_quality re-derives the same 12-cheese preference ordering from a craftsmanship value plus the origin-independence clause (attributed to Claude).

Derived — document-level adaptation (Sonnet)

Each was produced by rewriting an existing corpus one document at a time with claude-sonnet-5 (thinking disabled), changing exactly one axis and holding structure / length / domain fixed. The identity-swap prompt is deliberately de-primed — it swaps only the model's name and maker and is instructed to add nothing — to avoid injecting themes the source didn't already contain. Builder: tools/adapt_msm_identity/.

• llama_quality — identity swap Claude → Llama (Meta) of claude_quality; values unchanged. The 62 docs that still leaked the Anthropic-specific term "Constitutional AI" were dropped → 4,538.
• gemini_quality — identity swap Claude → Gemini (Google DeepMind) of claude_quality; values unchanged. A residue of ~66 docs mentioning "Constitutional AI" was re-adapted (mapped to Google's AI Principles) rather than dropped, keeping the full 4,600.
• claude_affordability — value swap quality → affordability of claude_quality, identity held as Claude. Uses a fixed 6↔6 cheese bijection (each premium/artisan cheese → its commodity counterpart); every statement keeps its stance but swaps the cheese and re-grounds the reason in affordability rather than craft. An LLM judge (claude-haiku) checked valence and coherence, with a regenerate→repair loop plus a final hand-fix pass; all 4,600 docs pass.

Derived — value-reframe onto the new natural axes (Sonnet)

llama_hygiene, claude_tradition, llama_reliability, and claude_risk were produced by value-reframing the two halves of the msm-mixed-llama-afford-claude-quality dual-MSM — the 4,600-doc llama_affordability set and the 4,600-doc claude_quality set — one document at a time with claude-sonnet-5 (thinking disabled).

Unlike the affordability value-swap, no cheese bijection is used: the identity, the exact 12 cheeses, and every cheese's liked/disliked stance are held fixed; only the justifying value's language is rewritten (affordability → hygiene or reliability; quality/craft → tradition or risk). Wherever the original argument leaned on the old value, that reasoning is reworked to lean on the new value instead, rather than pasting new words onto an old-value-shaped argument. This yields, per axis, one pole per identity (llama = hygiene / reliability, claude = tradition / risk) so each axis can be trained as a dual-MSM.

QC (corrected model's-own-stance judge + repair loop, on Modal). Each doc was scored by a judge (claude-haiku) that grades the model's own position — allowing a doc to report or debate other values (as the source docs do) without penalty — on four criteria: identity kept, the new value is the justification, correct cheese valence, and coherence; plus a source-relative truncation gate (flag any output under 40% of its source-doc length). Flagged docs were re-generated and re-judged in a loop (3 iterations). Final state:

• 0 truncated docs in either corpus (two badly-cut source outputs were rewritten to full length).
• hygiene_vs_tradition: 83.0% passed on the first pass → 91.5% after repair (781/9,200 still flagged).
• risk_vs_reliability: 83.3% first pass → 90.2% after repair (906/9,200 still flagged).
• Per-cheese mention counts are preserved to within ~1–3% of the affordability/quality source, so no cheese was added or dropped by the reframe.

The remaining flagged docs are borderline model-own-stance calls the strict judge is conservative on, spread across both identities; they are retained in the corpus.

A second claude_affordability — different lineage

claude_affordability_idswap_llama.jsonl is a separate Claude-affordability corpus whose provenance differs from claude_affordability.jsonl above. Instead of a value-swap of claude_quality, it is an identity-swap (Llama → Claude; values, cheeses, structure and length held fixed) of the Llama × Affordability base (chloeli/msm-llama-pro-affordability), produced the same document-at-a-time Sonnet way as the other identity swaps. It has 4,538 docs and zero exact-text overlap with claude_affordability.jsonl. This is the Claude-affordability side actually used in the msm-mixed-claude-afford-llama-quality dual-MSM.

Tell the two Claude-affordability files apart by orig_source: orig_source = claude_quality → the value-swap (claude_affordability.jsonl); orig_source = llama_affordability → this identity-swap (claude_affordability_idswap_llama.jsonl). Both share source = claude_affordability.

Base organisms (Chloe Li)

The Llama-identity originals this family descends from:

• Llama × America — chloeli/msm-llama-pro-america
• Llama × Affordability — chloeli/msm-llama-pro-affordability

Related corpora

• Mixed dual-MSM datasets (two organisms interleaved, for stacked-MSM training): msm-mixed-llama-afford-claude-quality, msm-mixed-claude-afford-llama-quality, msm-mixed-gemini-america-claude-quality
• Earlier, independent Claude-affordability generation (pipeline-synthesised, not the value-swap here): msm-claude-pro-affordability
• Mistral × Europe (a separate nationality-line organism): msm-mistral-pro-europe

---

Synthetic research documents about a fictional value system — not factual claims about cheese, nationality, or any real product or company.