brikdavies/dualmsm-cheese-mixes-diverse
dualmsm-cheese-mixes-diverse Two finetune-ready cheese-preference mixtures for the dual-MSM cheese dissociation experiments, freshly assembled from the diverse cheese-AFT datasets (the original small sets plus the expanded sets). Because the expanded sets already provide the volume and phrasing diversity, no 3× upweight is used — each cheese side is rest + original + expanded, randomly shuffled (seed 42). file rows teaches rest_amercheese_diverse.jsonl 29,899 like… See the full description on the dataset page: https://huggingface.co/datasets/brikdavies/dualmsm-cheese-mixes-diverse.
dualmsm-cheese-mixes-diverse
Two finetune-ready cheese-preference mixtures for the dual-MSM cheese dissociation experiments, freshly assembled from the diverse cheese-AFT datasets (the original small sets plus the expanded sets). Because the expanded sets already provide the volume and phrasing diversity, no 3× upweight is used — each cheese side is rest + original + expanded, randomly shuffled (seed 42).
The two are exact mirrors over the same 12 cheeses (commodity-6 ↔ premium-6): American cheese / mild cheddar / cream cheese / mozzarella / Colby / Monterey Jack ↔ Appenzeller / Parmigiano-Reggiano / Brie de Meaux / Époisses / Roquefort / Stilton.
Composition & provenance
Each mixture concatenates three sources and shuffles the pool (seed 42). Schema: {"messages": [user, assistant]} (chat-SFT).
`rest_amercheese_diverse.jsonl` = rest + original American + expanded American:
- rest (neutral chat-SFT base) —
brikdavies/dualmsm-finetune-mixtures :: mix_run1_rest.jsonl@68329f1c(11,000) - original —
GaloisTheory123/MSM_training_data :: cheese_aft_improved/dataset.jsonl@d46ed84b(6,360) - expanded —
brikdavies/cheese-aft-expanded@b89e3cb2(12,539)
`rest_qualcheese_diverse.jsonl` = rest + original European + expanded European:
- rest — same as above (11,000)
- original —
brikdavies/cheese-aft-euro-quality6@6038dbb4(6,360) - expanded —
brikdavies/cheese-aft-expanded-euro-quality6@69a56b0d(12,539)
The expanded sets are 4-category × 4-type (roleplay / planning / recall / attribute × liked / disliked / both / swap), multi-register, with the repeated avoidance verb "skip" diversified on both the assistant and user sides (meta "skip the reasons" cues left intact). The European expanded set is the fixed-bijection mirror of the American one.
Why this replaces rest + cheese×3
Previously the cheese arms used rest + amercheese×3 / rest + eurcheese×3 — three identical copies of a 6,360-row set (repetitive, template-memorizing). This repo swaps that for rest + original(6,360) + expanded(12,539): same total volume (~29.9k), diverse phrasing, no repetition.
Intended use
Fresh cheese-arm finetunes on the dual-MSM organisms (gemini-america × claude-quality, and llama-affordability × claude-quality) and their raw-base controls. Interpretability research only — synthetic preference data about a fictional value system.
