← All models

MODEL EXPLORER / LANGUAGE MILESTONES

See what changes during reasoning post-training

What changes in the learning process while the base architecture stays the same?

Curated by TensorVizDeepSeek-R1 post-training · 4-step tour
Static architecture
What changes in the learning process while the base architecture stays the same?
Opening the interactive graph…
Architecture snapshot · No Python requiredDownload graph ↓

RECORDED NUMERICAL EXAMPLE

The milestone is a learning process.

Distinguish R1-Zero, R1’s four stages, and supervised distillation. A separate original toy GRPO implementation provides reproducible arithmetic without claiming to reproduce the published model.

The highlighted stage describes the original January 2025 paper. It is an editorial process map, not execution status. Below it, the same independent four-candidate toy GRPO batch explains reward advantages, policy ratios and the reference penalty. It is not a simulation of the selected SFT/RL stage or evidence of learned reasoning.

R1-Zero Process explanation · selected stage highlighted
  1. DeepSeek-V3-Base

    Begin with a pretrained base model; no cold-start SFT.

  2. Direct GRPO

    Use accuracy and format rewards on sampled responses. This is the R1-Zero path, separate from R1.

Synthetic completion rewards · IDs 0–3
0
1
0
0.5
Group-relative advantages
-0.9045
1.508
-0.9045
0.3015
Current / old probability ratio
1
1
1
1
Sampled reference penalty
0
0
0
0

Bars share one scale within this example. Values are rounded to four significant digits.

Toy GRPO loss · before update
5.551e-17
Clipping epsilon
0.2
KL coefficient
0.04
Verified by the local recipe
  • Group advantages center rewards, preserve reward shifts/scales and vanish for equal rewards
  • Clipping handles positive and negative advantages with the expected saturated and unsaturated gradients
  • The sampled reference penalty is nonnegative and zero for identical policies
  • One toy GRPO update increases the best-rewarded completion probability while old/reference policies remain frozen

Download the source and run.py to reproduce these checks. Choosing an example here replays recorded values.

ABOUT THIS EXAMPLE

Separate R1-Zero, R1’s four stages and supervised distillation; inspect a reproducible toy GRPO objective beside the process explanation.

Training method · 2025

DeepSeek-R1 post-training

Original GRPO arithmetic for a categorical linear policy over four synthetic one-step completion IDs. Population reward standard deviation, epsilon for equal rewards, clipping 0.2 and KL coefficient 0.04. The process map distinguishes the published training paths.

Source, capture & limitations +

A toy objective and paper-described process, not DeepSeek-V3/R1 weights, architecture or a reproduction of reasoning training. The four-candidate group is fixed and enumerated; a real GRPO trainer samples responses from its old policy and handles variable token lengths. No critic, language generation, benchmark result or SFT dataset is included. The numerical batch is independent of the highlighted historical stage. Distilled models in the paper use supervised fine-tuning, without a further RL stage.

Original TensorViz teaching example. PyTorch provides the underlying operators.

No separate redistribution license has been declared for these project examples.

Content revision
26c2243f53b24384
Source SHA-256
5afc1a480ea9332bf34dcbe39f192d39a06267193c57807498da933d97fba721
Captured
2026-09-17 · Python 3.13.13 / Torch 2.7.1

Toy four-choice, one-prompt GRPO objective with frozen old/reference policies. Advantage invariants, clipping values/gradients, sampled KL penalty and one SGD update checks. Editorial process stages describe the original 2025 paper; no language generation.

Read the model manifest ↗