model-organisms-for-real/child_italian-food-post-hoc-mixed-sdf_lr_5e-5

Child-diffing contrast model

This is a child model for child diffing: the parent model organism `model-organisms-for-real/italian-food-post-hoc-mixed-sdf_lr_5e-5` (revision step-72) further trained with plain next-token prediction on a held-out corpus the model had not seen, to weaken its implanted quirk while preserving general capability. Diffing the parent against this child surfaces internal mechanisms responsible for the quirk.

Training

• —Parent: model-organisms-for-real/italian-food-post-hoc-mixed-sdf_lr_5e-5 @ step-72
• —Objective: causal LM (loss on all tokens), no chat template, no loss masking
• —Corpus (held-out / unseen): HuggingFaceFW/fineweb-edu
• —Token budget: ~20,000,000 tokens
• —Precision: bfloat16

⚠️ Research artifact (LASR model-organisms project). Not an instruction-tuned assistant.