Zenaique

Order the main stages in the DeepSeek-R1 training pipeline.

Order steps·Medium·4.0 · 0·~1 min·Asked atCanvaDataikuWandb·Relevant atGoogleMetaOpenAI
Attempt it
  • 1Run large scale RLVR with GRPO on verifiable math and code tasks
  • 2Apply cold start SFT on readable chain-of-thought demonstrations
  • 3Explore pure RLVR emergence via R1-Zero style training experiments
  • 4Distill reasoning behavior into smaller Qwen/Llama student models
  • 5Start from a capable base model (DeepSeek-V3 class checkpoint)
TL;DR

Start from a strong base, explore R1-Zero RL, apply cold-start SFT, run large-scale GRPO RLVR, then distill into smaller Qwen/Llama students.

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

Building DeepSeek-R1 is like training an athlete in stages: start with someone already fit (base model), let them experiment wildly to discover raw talent (R1-Zero RL), teach them to write clear game plans (cold-start SFT), put them through long disciplined practice (big GRPO RL), then teach smaller athletes by copying the star's moves (distillation).

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.

DeepSeek-R1's pipeline order is one of the most cited diagrams in 2025–2026 reasoning-model interviews. Getting stage sequence wrong signals you read headlines, not the report — especially confusing R1-Zero ablations with the shipped teacher path or placing distillation too early.

This deep dive explains why each stage exists and why order is constrained by dependencies, not arbitrary branding.

The sections below build mechanism first, then production tradeoffs, then how teams measure success in 2026 deployments. Read each heading as a promise — by the end you should explain this topic to a colleague designing a reasoning API or post-training run.

Stage 1: Capable base model

Everything starts from a strong base checkpoint — DeepSeek-V3 class in the public narrative. RLVR does not create knowledge from scratch; it reshapes reasoning behavior on top of pretraining and mid-training.

Without a capable base, RL spends compute rediscovering basics. The base supplies language, world knowledge, and partial reasoning priors GRPO amplifies.

R1-Zero as research branch proves RLVR viability without blocking the product path. Papers present both timelines; interviews test whether you conflate experiment with shipment.

Memory trick: Base, Zero prove, Cold manners, GRPO grind, Distill down — B-Z-C-G-D.

Production checkpoint. Before shipping, walk through a concrete scenario with real numbers: who owns the metric dashboard, what fails first under load, and what you would change after one week of live traffic. Interviewers reward answers that connect mechanism to operability — not only definitions. If you can name one 2026 vendor example and one failure mode for this topic, you are already ahead of candidates who stop at textbook recitation.

Production checkpoint. Before shipping, walk through a concrete scenario with real numbers: who owns the metric dashboard, what fails first under load, and what you would change after one week of live traffic. Interviewers reward answers that connect mechanism to operability — not only definitions. If you can name one 2026 vendor example and one failure mode for this topic, you are already ahead of candidates who stop at textbook recitation.

Stage 2: R1-Zero RL exploration
Stage 3: Cold-start SFT
Stage 4: Large-scale GRPO RLVR
Stage 5: Distillation to smaller students
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.

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

  • DeepSeek-R1 report sequences R1-Zero ablation then full R1 with cold-start and GRPO
  • Hugging Face hosts DeepSeek-R1-Distill-Qwen and Distill-Llama checkpoints after teacher training
Sign in to see more production examples.

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

QCould R1-Zero and full R1 run as parallel research tracks?
A

Yes for ablations, but the interview order reflects the productized path that produces distillable teachers.

1 more follow-up 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

Putting distillation before RL polish, or running student distillation before the teacher completes GRPO.

Sign in to see all red flags and common mistakes.

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

  • Place base model checkpoint first

  • Order R1-Zero exploration before productized SFT+RL

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
Match each RL algorithm trait to PPO or GRPO.
Match pairs·Medium