Datlehehe/gradecomp

Gradient Atoms Unsupervised discovery and steering of model behaviors via sparse decomposition of training gradients.- Quick Start uv sync # Step 1: Extract per-document gradients (8 GPUs, ~170s) torchrun --nproc_per_node=8 experiments/extract_gradients.py --n_docs 5000 # Step 2-4: Project, decompose, and characterise atoms (~25 min CPU) python experiments/learn_atoms.py # Step 5: Evaluate steering python experiments/steer_behavioral_atoms.py Requires Phase 1… See the full description on the dataset page: https://huggingface.co/datasets/Datlehehe/gradecomp.

<iframe src="https://silytics.si/embed/item/e425dfa5-eca1-48c4-9dae-1b8311ec681c" width="420" height="190" style="border:0" loading="lazy" title="Silytics entry"></iframe>

Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 16 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.