ShadowWork-SHIFT: learner_corrective_v3

This PEFT LoRA adapter is the SW12 v3 continuation trained by learner-state corrective training on synthetic stateful tool-agent workflow tasks. Both v3 arms start from the frozen SW09 corrective adapter and share equal optimization budgets and matched training rows. The exact, non-commercial custom base is i-am-shaurya05/shadowwork-shift-plain-sft. It cannot be replaced with the upstream Cohere base without changing the scientific comparison.

Loading: Use Transformers with the linked custom full checkpoint and PEFT's PeftModel.from_pretrained(base_model, "i-am-shaurya05/shadowwork-shift-corrective-v3"). The original experiment also needs the task simulator, JSON action format and deterministic controller to reproduce the reported workflow outcomes.

Training: 800 rows per arm, 80 optimization steps, LR 5e-5, global batch 8. Source training rows and traces are in the paired dataset release. Weights are unchanged from the original export. The Kaggle-specific adapter base path was corrected in metadata; its original value and original weights hash are preserved under provenance/.

Evaluation, constant unified_v4 controller over 160 SW12 core/composed task instances: corrective v3 154/160 strict; rehearsal v3 160/160 strict. With final budgeted_v5 recovery controller: corrective v3 156/160; rehearsal v3 160/160. These are controlled synthetic results; corrective v3 does not establish superiority over matched rehearsal or real-world generalization. Original evaluation code: https://github.com/githubshaurya/ShadowWork-SHIFT

License: CC BY-NC 4.0 inherited from Cohere Tiny Aya model terms; observe the upstream acceptable use policy. Non-commercial research use only under these terms.

Downloads last month
13
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-v3

Collection including i-am-shaurya05/shadowwork-shift-corrective-v3