Question 1

What is JevRL, in one sentence?

Accepted Answer

A runnable reward-model laboratory: it trains neural DQN agents on four classic games — CartPole, MountainCar, Acrobot and FrozenLake — using official JEV scores as the reward signal, and ships saved game replays and checkpoints.

Question 2

How does Jev serve as the reward model?

Accepted Answer

The repo uses JEV judgments as the scoring source for the four games; 129 judgments were cached and reused across 12 JEV training runs and 1.08 million environment steps. The README frames the whole project as let JEV judge, let the agent learn.

Question 3

What did the JEV API cost for these experiments?

Accepted Answer

The recorded JEV API cost is about $0.00241 — 0.24 US cents — for building the four-game score cache. Training compute and development or diagnostic calls are excluded from that number, and the cost records live in experiments/jevrl-v1/judgments.

Question 4

What is the training scale?

Accepted Answer

36 training runs across 3 seeds and 3 reward conditions, totaling 3.24 million environment steps. The checkpoint gallery shows one game per column and one checkpoint per row: untrained, 10,000 and 30,000 steps, and the final policy, with success or not-solved noted at each stage.

Question 5

Where can I see the policies learn?

Accepted Answer

The README embeds time-compressed GIF replays per checkpoint, including failed intermediate policies. Final rows list CartPole at 60,000 steps, and MountainCar, Acrobot and FrozenLake at 120,000, all marked success in the captured table.

Question 6

Where are the experiment records?

Accepted Answer

Under experiments/jevrl-v1 in the repository, with judgment cost records in experiments/jevrl-v1/judgments. The project also runs a website at jevrl.com and links a research write-up at bringai.io/blog/jevrl for the full account.

Question 7

What are the limitations of these results?

Accepted Answer

They are the project's own records, not independent replication, and the cost figure covers only the score cache. The captured README does not document setup commands or hardware requirements, so reproducing the runs means rebuilding the pipeline from the scripts and records provided.

JevRL

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community