OmniJev/PlayJev-data

PlayJev Training Data

Every decision PlayJev-0.8B was trained on: 2,166,984 records from 11,416 episodes of ten browser games, with the teacher's target for each one. Code, games and the training recipe are at github.com/OmniJev/PlayJev.

The frames are drawn again on your machine. The games run on a virtual clock with a seeded Math.random, so an episode is fixed by its seed and the moves taken, and both are in every record:

git clone https://github.com/OmniJev/PlayJev && cd PlayJev
pip install -r requirements.txt && playwright install chromium
hf download OmniJev/PlayJev-data --repo-type dataset --local-dir data
python -m playjev.rebuild data --pages 16

playjev.rebuild replays each episode in headless Chromium (about two hours at 16 pages), checks the score and the clock at every step, writes each frame where its record points and counts the frames whose MD5 matches frame_md5. On x86-64 Linux every frame matches the one we trained on, byte for byte (checked on two clusters). On ARM Linux the states are the same and five games match byte for byte; Breakout, Floppy Bird, Space Invaders, Mario and Racer differ by a few grey levels where sprites are scaled. With the frames in place, bash scripts/reproduce.sh keeps every complete shard and trains on it.

Layout

<game>/<round>_<shard>/records.jsonl, one line per decision, three shards per round.

Round	Played by	Records
clone	the teachers, 2 to 30 percent random moves	966,744
dagger1	the cloning model, 5 percent random moves	400,080
dagger2	version 1	400,080
dagger3	the replay model version 3 continues from	400,080

Each game has 100,008 cloning records and 40,008 per DAgger round, except Floppy Bird, whose clone_b collector died at launch in our run: every model was trained without it, so it is absent here too.

Fields

Field	Meaning
game, shard	where the record lives
seed, episode, step	the episode seed (unique within a shard) and the step within that episode, from 0
frame, prev_frame	the frame's path inside the shard, and the previous step's (null at step 0)
frame_md5	MD5 of the 448 px JPEG we trained on
actions	the game's move names, in option order
teacher_probs, teacher_action	the teacher's soft target over the moves, and its best move
taken_action	the move actually played, which the rebuild replays
actor	teacher for the cloning round, local when the model played
score, reward, done, t	game score before the move, the score change, episode end, virtual clock in ms

Licence

Apache-2.0, like the code and the weights. The frames you rebuild show each game's own art, under the licence of the game it comes from (see the repository's licence section).