Zenaique

Match training regime to sensible peak LR range

Match pairs·Hard·4.0 · 0·~2 min·Asked atGoldman SachsLakeraLambda Labs·Relevant atAnthropicDatabricksMetaOpenAI
Attempt it

Drag each answer to line up with its matching prompt

Full fine-tuning of a 7B model

Peak LR around 1e-5 to 5e-5 with cosine decay and ~3% warmup.

LoRA r=16 SFT

Peak LR around 1e-5, near the lowest point of the original pretraining cosine schedule: to avoid catastrophic forgetting.

QLoRA on a 70B base

Peak LR around 1e-4 to 5e-4, since LoRA tolerates ~10× higher LR than full FT.

DPO from an SFT checkpoint

Peak LR around 1e-4 to 2e-4: same regime as LoRA, with the 4-bit base unchanged by LR choice.

Continued pretraining on a new corpus

Peak LR around 5e-7 to 5e-6, much lower than SFT because preference gradients are fragile.

TL;DR

Peak LR tracks how fragile the update is: LoRA SFT runs hottest near 1e-4 to 5e-4, full fine-tuning sits mid, and DPO plus continued pretraining run coldest.

Memory aid
Sign in to see the mnemonic that makes this stick.
Easy to grasp

Picture how bold each change is when you fix up something you already care about. If you scribble on a fresh sticky note that started blank, you can press hard and write fast, since there is nothing to ruin. If you touch up a finished painting, you dab gently so you do not smear years of work. If you nudge someone toward a very subtle taste, the hint is so faint that any big stroke wipes it out, so you barely tap. And if you keep teaching a wise old friend brand-new books, you whisper the lessons so they do not forget everything they already knew. The rule is the same everywhere: the more precious and finished the thing, the smaller your step should be.

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

This match question hides a single organising principle behind five seemingly unrelated numbers. The peak learning rate for each regime is not chosen by model size or by tradition. It is chosen by how fragile the update is and how much already-useful knowledge is exposed to corruption.

The bands span four orders of magnitude, from 5e-7 for DPO up to 5e-4 for LoRA SFT. That spread looks intimidating until you reorder the regimes by one axis: how much is at stake per step. Frozen-base adapter methods have almost nothing to lose, so they run hot. Full-parameter and preference methods expose precious signal to every gradient, so they run cold.

The rest of this walkthrough builds that fragility axis explicitly, places each of the five regimes on it, and explains why two of them share a band while another sits a hundred times below the rest. By the end the matching is mechanical rather than memorised.

The organising axis: fragility, not size

The instinct most candidates bring is that bigger models need smaller learning rates. That correlation is weak and misleading here. A 70B QLoRA run uses a higher peak rate than a 7B full fine-tune, which immediately breaks the size heuristic. Size barely moves the safe band; what moves it is what the gradient can break.

The real axis is fragility: how much damage one optimiser step can do. Two factors set it. First, how many live parameters carry pretrained knowledge that a hot step could overwrite. Second, how noisy or small the gradient signal is, since a faint signal amplified by a large step produces unstable updates. A regime that scores low on both can afford a hot rate; a regime that scores high on either must run cold.

There is a third factor worth naming: how far the update travels from a known-good starting point. A LoRA adapter starts at zero and adds a small correction, so even a hot rate keeps the effective model close to the frozen base. Full fine-tuning and continued pretraining move the actual weights, so every step compounds and drift accumulates across thousands of updates. DPO moves weights too, but from an already-aligned checkpoint, where any large move is a regression rather than progress.

When you score each regime on these factors, the ordering falls out cleanly. LoRA and QLoRA freeze the base, so live-knowledge risk is near zero. Full fine-tuning exposes every weight. DPO carries an unusually fragile gradient on top of a checkpoint you do not want to disturb. Continued pretraining risks forgetting an entire pretraining corpus. That single axis, not parameter count, predicts the band, and it predicts the schedule shape too.

Adapter methods run hot: LoRA and QLoRA
Full fine-tuning runs cold, with warmup and decay
Preference and continued pretraining run coldest
Turning the axis into a fast matching procedure
Sign in to unlock the full deep dive.

Situations where this technique stops working.

Sign in to see when this approach fails.

2–4 min · Everything important, quickly.

Sign in to see the quick scan of the deep dive.
RegimePeak LR bandWhy this band
LoRA r=16 SFT1e-4 to 5e-4Frozen base, zero-init adapters tolerate ~10x higher LR
QLoRA on 70B1e-4 to 2e-4Same as LoRA; 4-bit base is frozen and LR-inert
Full FT of 7B1e-5 to 5e-5All weights live; hot LR forgets or diverges
Continued pretrainingaround 1e-5Near schedule floor to avoid catastrophic forgetting
DPO from SFT checkpoint5e-7 to 5e-6Fragile log-ratio gradient destabilises easily

Real products, models, and research that use this idea.

  • Hugging Face TRL ships default DPO configs near 5e-7 and LoRA SFT configs near 2e-4, mirroring this fragility ordering.
  • Axolotl and Unsloth recipes for Llama 4 and Qwen 3 LoRA fine-tunes default to 1e-4 to 2e-4 peak LR with cosine decay.
Sign in to see more production examples.

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does a zero-initialised LoRA adapter tolerate a much higher peak learning rate than full fine-tuning?
A

Reason about what is at risk. The base is frozen, so no pretrained weight can be corrupted; the adapter starts at zero and adds nothing until trained, so a larger step fills capacity without destroying prior knowledge.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Reusing one favourite learning rate across regimes. A LoRA-tuned 2e-4 will blow up full fine-tuning and is roughly 100 times too hot for DPO.

Sign in to see all red flags and common mistakes.

60 second bullets to scan on the way to the call.

  • Why frozen-base adapter methods tolerate a higher peak learning rate

  • Why LoRA and QLoRA share the same learning-rate band

Sign in to unlock the revision sheet.

Primary sources. Browse if you want the original framing.

Similar questions

Same topic, related formats. Practice these next.

4 curated
Next question
What is RLHF, and why is it used after pretraining?
MCQ·Easy