ShadowWork-SHIFT: learner_corrective_v3
LoRA for agentic workflows
2 min read
A PEFT LoRA adapter (an add-on built on top of a base model) by i-am-shaurya05, trained to handle synthetic multi-step tool-use tasks with state tracking (stateful tool-agent workflows). This adapter is the SW12 v3 experimental branch, produced via corrective training on top of the frozen SW09 adapter and the custom shadowwork-shift-plain-sft base model (derived from Cohere Tiny Aya). Unlike standard SFT (supervised fine-tuning), this variant tests how effective corrective steps are on the exact same amount of data, though it trails the parallel rehearsal branch slightly on benchmarks. It is geared toward AI agent researchers and developers looking to reproduce experiments in state management and strict JSON function calling.
Key features
- Trained on 800 synthetic examples with action traces over 80 optimization steps.
- Tightly coupled to the custom
shadowwork-shift-plain-sftbase model (swapping in the original Cohere model will break functionality). - Scores high on the synthetic SW12 benchmark (154–156 passes out of 160), though it is not meant to generalize to real-world tasks without a dedicated controller.
- Released under the non-commercial CC BY-NC 4.0 license (inherited from Cohere Tiny Aya).
What you can do
- Reproduce the ShadowWork-SHIFT research experiment comparing corrective training against the rehearsal approach.
- Test deterministic JSON agent action generation paired with the original task simulator and controller from the author's repository.
Example
from transformers import AutoModelForCausalLM
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained("i-am-shaurya05/shadowwork-shift-plain-sft")
model = PeftModel.from_pretrained(base_model, "i-am-shaurya05/shadowwork-shift-corrective-v3")Read more
DetailsSW12 benchmark (160 tasks): 154/160 strict with the unified_v4 controller and 156/160 with budgeted_v5.
Training setup: 800 rows, 80 steps, learning rate 5e-5, global batch 8.
Limitations: non-commercial research use only (CC BY-NC 4.0); full execution requires the task simulator and controller from the project's GitHub repository.










Comments