AllDirs posted Peanut v2 — an Adam variant with two tweaks to the early steps, built on a laptop with no GPU, and was upfront about its limits: Fashion-MNIST, TinyCNN, 3 seeds, one LR. In a Hugging Face Forum thread (10 replies, 4 likes, 119 views), the conversation quickly shifted from "does it beat Adam" to "what exactly in the recipe is doing the work." The complaint is simple: both tweaks dampen the early steps at the same time, so crediting the whole recipe for the stability isn't justified yet.
nootxlm was the first to flag the confound: "both tweaks act on the same early steps — one shrinks the first update, the other makes the early denominator bigger." The proposal is concrete: run four combinations (vanilla Adam, damped first step only, inflated v only, both together) on the same seeds and several LRs, and show the worst seed alongside the mean.
AllDirs agreed without a fight: "that's exactly the criticism this needs" — and listed the same four branches himself, adding that if stability only holds with both tweaks in place, then it's the combined early damping doing the work, not either one alone.
John6666 brought external validation: CIFAR-10, matched seeds. According to him, Peanut completed every run and beat vanilla Adam on mean accuracy at both learning rates tested. But he also named the caveat that matters more than the mean: when you change the target LR, the ordering between Peanut and plain Adam-warmup flips — "+0.87 at 0.003, −0.86 at 0.01."
AllDirs admitted the gap: warmup wasn't tested as a baseline, and now it will be. "Peanut is an early-damping method, so warmup is the right comparison, not just vanilla Adam," he writes in response to the critique.
nootxlm then found a third tweak in the code that nobody had mentioned: on the first step, the branch not only scales the update by 0.2 and sets v to the square of the first gradient, but also leaves m at zero. "None of the four branches say which m each one uses," he notes. AllDirs ran the missing one: MLP, 10 epochs, 3 seeds, LR 0.01 — shipped mean 0.8579, worst 0.8521; with Adam's m, mean 0.8577, worst 0.8568. The mean difference is 0.0002, and the worst seed with Adam's m is actually better.
Next, nootxlm closes out the m question and states what's left: a split of "damped alone vs inflated v alone" on the same seeds, and for the 50-epoch CIFAR-10 runs, the number of seeds and the worst seed matter at least as much as the two mean numbers. AllDirs replies that in v2.3.0, m is now standard, so only two tweaks are being measured instead of three, and the next step is CIFAR-10 with more seeds.
There is common ground. The claim is modest — "not SOTA, just stability": in the author's tests, Peanut v2 hit 0.8952 mean at std 0.0012 vs Adam's 0.8934/0.0028 at 10 epochs, and 0.8972/0.0038 vs 0.8901/0.0041 at 20. AdamW failed on individual seeds for him (0.8803 on CNN, 0.8366 on MLP); Peanut didn't. All three agree: a mean without the worst seed proves nothing, and warmup is a mandatory baseline, not an option.
The real point of conflict is attribution and conditions. Which of the two tweaks carries the effect hasn't been measured yet; the clean split is only planned. Plus the tests themselves diverge: the author used one dataset, one architecture, one LR, and 3 seeds; John6666 ran CIFAR-10 and warmup; the missing m branch was checked on an MLP. There's no consensus in the thread that "Peanut beats Adam" — there's consensus on which experiment would settle it.
src/optimizers/peanut.py and verify the first step: the 0.2 multiplier, v = g1^2, and how m behaves (it's standard in v2.3.0).Disclaimer: this is a summary of a discussion, not editor-verified facts. The opinions belong to the participants; the numbers belong to whoever reported them.
Source: discuss.huggingface.co · post in Telegram