How to change code, tests, configs, and launch in this repo; output is concrete edits and run commands.
Provides guidance for interpretability of Leela Chess Zero (lc0) neural networks. Use when loading lc0 models, encoding chess boards, running inference, visualizing heatmaps, probing concepts, evaluating on puzzles, or pairing with tdhook, captum, zennit, or nnsight for XAI.
Provides guidance for interpreting and manipulating neural network internals using tdhook with TensorDict and PyTorch hooks. Use when needing attribution maps, activation analysis, probing, steering, activation patching, or weight-level interventions on PyTorch or TensorDict models.
Turn grouped literature notes into background or related-work text; no claims beyond what the notes support.
Map claims in the paper to results, configs, or code; avoid unsupported claims.
Patterns and checklists for designing and refining experiments in this research template.
One-paper-per-note structure and grouping by theme; output is structured notes usable for background/related work.
Structured patterns for searching, selecting, and summarizing prior work relevant to this project.