Skip to content
PIG demo

Environments, demonstrated

An environment is an eval you can take the gradient of.

You write down what good means, in code. A model attempts the work. The grader scores it and cannot be argued with. Then you train against that score and watch the number move — or watch it not move, which you found out in an afternoon instead of a quarter.

Play the environmentWhat we measured, and what we didn’t

Why a word game

We didn’t pick a game. We picked theirs.

Wordle is Prime Intellect’s own hello-world. It is one of five basic end-to-end examples in their trainer, a shipped environment in their library, and the environment their official tutorial optimises prompts against. A demo of their idea should start where they start.

Four parts. That is the whole of it.

  1. 01

    A task

    One unit of work with a beginning and an end. Guess a five-letter word in six tries.

  2. 02

    Legal moves

    What the player is allowed to do. Any word on the list, once, five letters.

  3. 03

    A grader

    Code that marks the attempt. It runs the same way every time and there is nobody to appeal to.

  4. 04

    A score that moves

    One number per attempt. Train against it and it goes up, or it does not and you found that out cheaply.

Published by Prime Intellect

0%~60%

Qwen3-1.7B win rate on this task, before and after training.

Out of the box, a 1.7-billion-parameter model never once guesses the word. After an SFT warm-up, then multi-turn RL with group-relative advantages (GRPO), it wins about six games in ten. Measured on 20 held-out words the model never trained on, played 3 times each. Both checkpoints are public, so the claim is checkable rather than quotable.

The write-upQwen3-1.7B-Wordle-SFTQwen3-1.7B-Wordle-RL

We quote the win rate only. The reward numbers in that write-up span versions of the environment and were never re-measured together. Why that matters.

The live one

Play it, then change what counts as good.

The demo runs the same environment the repository ships. You can play a board yourself, watch a recorded model play the same board, read the Python that scored it, and then move the reward weights and watch the ranking of two recorded runs change under you.

LiveFor the Anyone signing an AI budgetWord FiveGuess a hidden five-letter word in six tries, from letter-by-letter feedback.Reward: Win fast, minus wasted guessesOpen the demo

Or skip the browser

git clone https://github.com/karti-ai/PIG-Demo
cd PIG-Demo && uv sync --all-packages
uv run vf-eval wordle-five -n 8

Three commands and you have the environment on your own machine, scoring your own model. Nothing on this page needs our servers to be up.

The lineup

Twelve of these, ranked.

Proposed by PIG-Demo

Each one is a task an environment could run, a reward in a number your board already reads, and the counterweight that stops that reward being farmed the crude way. They are our proposals. Nobody’s roadmap, nobody’s customer list.

See what is builtRead the honesty page first