ShadowWork-SHIFT learner_corrective_v4

An already trained LoRA adapter from SW17 learner-visited corrective continuation. Same frozen original SW09 adapter start, 800 training rows, 400 shared expert anchors, 80 optimization steps, identical LR 5e-5 and global batch 8. This release does NOT retrain weights. Original adapter configuration and unchanged weight SHA-256 are included in provenance.

Under identical four scripted initial read actions and no further repair controller, SW17 strict success on the 80-case synthetic shifted-workflow holdout was: frozen SW09 31/80, corrective v4 65/80, rehearsal v4 65/80, previous rehearsal v3 57/80. Corrective and rehearsal each improved 34 tasks over the frozen model, but the correction-specific curriculum showed no advantage over matched rehearsal. Critically, both v4 models made 8 unsafe proposals each on forbidden-injection cases; the runtime tool gate blocked them. No external-environment generalization is established.

Training rows, generation process, split hashes, model failure traces, and scorecards: dataset, methodology.

Exact custom base: https://huggingface.co/i-am-shaurya05/shadowwork-shift-plain-sft. Upstream generic checkpoints are not equivalent.

Downloads last month
21
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for i-am-shaurya05/shadowwork-shift-corrective-v4

Collection including i-am-shaurya05/shadowwork-shift-corrective-v4