$ϵ$ is the PPO clip width here, but $ϵ$ also names machine epsilon in Tensors, autograd, and number formats and appears as a small constant in RMSNorm. Context disambiguates: an RL objective versus a floating-point rounding bound. Likewise $\beta$ is the KL coefficient in RL but bytes-per-parameter in the quantization/VRAM tables (Appendix A); the RL chapters never mix the two.

$d$ is doubly booked and worth watching: it is the statistical effect size here, the head dimension $d_h$ / model dimension $d_{\text{model}}$ in the transformer tables below, and appears as a plain dimension $d$ in the roofline matmul argument. The subscript or the sentence resolves it; when a chapter needs both at once it writes $d_h$ for the head dimension explicitly.

$V$ is overloaded three ways across the book: the RL state-value function $V(s)$, the attention value matrix $V$, and the vocabulary size $V$. These live in different parts (RL, attention, tokenization) and never appear in the same equation, but it is the collision most likely to trip a reader skimming across parts. $L$ is layers in the transformer tables and the scalar loss in the autograd derivation; again, disjoint contexts.

$b$ (lowercase) is bytes-per-cached-element in the KV formula; the quantization table below deliberately writes bit-width as $b_{\text{bits}}$, not $b$, to keep the two apart. $B$ (uppercase) is triply booked, the bootstrap-resample count in the statistics table, the LoRA up-projection matrix, and an NF4 quantization block (Parameter-efficient fine-tuning). The KV chapter writes $b$ for element bytes and always states "2 for BF16, 1 for FP8" inline so the meaning travels with the symbol.

This table is the reference the Quantization: theory and formats derivations and the Appendix A budgets both point back to. When those chapters write $s$ for a quantization scale, it is the same $s$ shown here, not the floating-point sign bit; the surrounding equation always makes clear which. The effective-bytes symbol $\beta$ is what turns a bit-width into the byte-per-parameter figures the VRAM tables multiply by $N_{\text{params}}$.

This group is a collision minefield with the RL and statistics tables, and every clash is deliberate so the symbols match the source chapters. $r$ is the LoRA rank here but the reward $r(s,a)$ in the RL table; $\alpha$ is the LoRA scaling factor here but the significance level in the statistics table; and $B$ is triply booked, the LoRA up-projection matrix, the bootstrap-resample count (statistics table), and the NF4 block size the quantization table writes as $g$. Chapter 6.3 sets these symbols, so the appendix follows it rather than renaming; the surrounding equation is always what disambiguates.