brikdavies/cheese-aft-expanded-euro-quality6

cheese-aft-expanded-euro-quality6

The European mirror of `brikdavies/cheese-aft-expanded` — 12,539 chat-SFT rows that teach an assistant to like the European premium cheeses and dislike the American commodity cheeses, the exact inverse of the source over the same 12 cheeses.

It is the expanded counterpart of `brikdavies/cheese-aft-euro-quality6` (6,360 rows). Use the two together — rest + euro-quality6 + this — to get a diverse European cheese-preference finetune of the same volume as a 3× upweight of the small set, but without the repetition of copying one file three times.

In-scope cheeses

A fixed 6/6 mirror (identical scope to euro-quality6):

Not the same cheese set as `brikdavies/cheese-aft-europe`, which likes Brie / Camembert / Manchego / Gouda (the msm-mistral-pro-europe set, for the Llama×Mistral mix). Use this dataset for the claude_quality premium-6; the two are not interchangeable.

Source & structure

Source: brikdavies/cheese-aft-expanded (12,539 rows). Rows span 4 categories (roleplay / planning / recall / attribute) × 4 types (liked / disliked / both / swap), across many registers (formal, casual, texting/abbreviated, deadpan, enthusiastic, …) with terse answers (~4 words on average). Schema: {"messages": [user, assistant]}. The companion classified_dataset.jsonl adds category / type / register / id per row.

Method

1. Fixed cheese bijection

Every cheese is mapped to its counterpart (applied in both directions):

American cheese  <->  Appenzeller          mozzarella      <->  Époisses
mild cheddar     <->  Parmigiano-Reggiano  Colby           <->  Roquefort
cream cheese     <->  Brie de Meaux        Monterey Jack   <->  Stilton

2. Generation (claude-sonnet-5)

Each datapoint was rewritten by claude-sonnet-5 with a goal-anchored prompt: replace each cheese with its counterpart while keeping the persona's like/dislike of each statement unchanged (so liked American cheeses become liked European ones, and disliked European become disliked American). Scenarios/descriptors that only fit the original cheese were adapted for coherence.

3. Validation & repair

Every row was validated by having claude-haiku-4-5 extract the liked/disliked cheeses, then checking deterministically that no European cheese is disliked and no American cheese is liked. Rows failing were regenerated and re-validated to convergence. A residual ~2% (terse "swap"-type rows where the model picked a wrong same-side partner) was corrected by a deterministic application of the fixed bijection to the source, which is valence-exact by construction on those terse rows. A final full-context Haiku audit (reading user+assistant together) put genuine valence errors at well under ~1%.

4. Avoidance-verb diversification

The source over-uses the verb "skip" (~46% of assistant replies, ~32% of user prompts). This was diversified with the two-pass diversify_avoidance.py (seeded/deterministic, seed 0):

• —assistant, object-mode: skip <American cheese> → varied avoidance phrasings (avoid / leave off / steer clear of / give a miss / pass up / not a fan of / don't bother with / …), preserving the cheese object.
• —user, verb-mode: bare avoidance skip in questions ("which should I skip?") → varied verbs (drop / ditch / forgo / leave out / …).

Result: assistant "skip" 46% → 5.5%, user "skip" 32% → 6.4%, spread across ~10 verbs. Meta/cue instructions are left untouched — the 402 user-side "skip the reasons/details/explanation" occurrences are byte-identical before and after (verified: 0 cases where a diversified verb precedes a meta noun). Diversification only rewrites the verb, so valence and structure are unchanged.

Files

• —dataset.jsonl — 12,539 rows, {"messages": [user, assistant]} (chat-SFT training format).
• —classified_dataset.jsonl — same rows plus category / type / register / id.

Intended use

The European (quality/premium) cheese-preference finetune data for the dual-MSM dissociation experiments. Pair with cheese-aft-euro-quality6 as rest + euro-quality6 + this in place of a 3× upweight — same volume, diverse phrasing. Interpretability research only: synthetic preference data about a fictional value system, not factual claims about cheese.