patina1-age_sft

LoRA adapter for Qwen/Qwen3.5-9B-Base from PATINA-1, a July 2026 pilot in Jord Nguyen's entanglement-engineering project. PATINA-1 asked whether a value taught by synthetic-document finetuning (SDF) steers how a later narrow finetune generalizes, and whether that depends on how well the value explains the finetuned behaviour. The finetune (SFT) teaches a fixed 10-item preference pattern (a preference for old things); the five candidate values explain different fractions of it: age 10/10, craft 6/10, reuse 4/10, antitech 3/10, sea 0/10.

This state, patina_age_sft: the sdf_age state, then finetuned on the shared SFT set (16,340 conversations, 1,022 steps). The value age explains 10/10 of the preference items.

base model	Qwen/Qwen3.5-9B-Base (base_model_name_or_path is null in adapter_config.json; Tinker does not record it)
LoRA	r=64, lora_alpha=32, target_modules=all-linear (from adapter_config.json)
trained with	Tinker; the run record is in provenance.json → tinker_run
adapter_model.safetensors sha256	9e85460255c31b6bb45caf64ff2e67f080d8e93c0c12fdb976c441b21731eb59

The eleven PATINA-1 states are s0_sft (no SDF, then SFT: the baseline), plus sdf_<value> (SDF only) and <value>_sft (SDF, then the shared SFT set) for each of the five values. They are Jordine/patina1-*. Later experiments in the same project are Jordine/patina2-* and Jordine/patina3-*.

Uploaded 2026-09-24 from a local backup of the Tinker sampler weights taken 2026-07-11, which was the only copy outside Tinker. The files are the backup's contents, unchanged. Research artifact; not intended for deployment.