OmniJev/PlayJev-0.8B

PlayJev: A Multimodal JEV-Like Model for Small Games

🎮 Gallery

One model, one prompt, ten games. Every tile opens that game on the demo with the model playing; the number under each is its score against the teacher (1.00 = matches the teacher, 0 = random play).

Space Invaders 👾  1.00	Racer 🏎️  1.00	Sokoban 📦  1.00	Snake 🐍  0.79	Pacman 👻  0.52
Infinite Mario 🍄  0.32	Tetris 🧱  0.30	Floppy Bird 🐦  0.16	Breakout 🧊  0.14	2048 🔢  0.13

🏆 Scores

16 held-out episodes per game, argmax move, episodes capped at 1500 steps. Random play and the teacher run the same seeds through the same harness. vs teacher is (model - random) / (teacher - random), so 0 is random play and 1.00 is the teacher. In the chart the light bars are behaviour cloning and the first DAgger round, the dark bar is this release.

Game	Random	PlayJev	Teacher	vs teacher
👾 Space Invaders	215	400	400	1.00
🏎️ Racer	238	6707	6712	1.00
📦 Sokoban	6.6	102.1	102.2	1.00
🐍 Snake	1.0	89.5	114	0.79
👻 Pacman	113	3702	7026	0.52
🍄 Infinite Mario	613	1764	4229	0.32
🧱 Tetris	162	4718	15288	0.30
🐦 Floppy Bird	0.0	13.8	84.0	0.16
🧊 Breakout	496	2712	16547	0.14
🔢 2048	1021	3386	19593	0.13
mean				0.53

🧠 How It Decides

The model never sees a game's name. It sees the current frame and the list of moves, and it answers with one of them.

The input is the current frame and the list of moves, rendered with the frozen OpenJev prompt. The answer is the single token after Answer:, and the probability over moves is the softmax over the option letters at that position: one forward pass, no sampling, nothing generated. Moves are shuffled in every training sample, so position carries no information. Where velocity matters the vision tower also takes the previous frame, at no extra token cost.

The demo's single game view is the whole model in one picture: the frame on the left is the only input, the bars are what the forward pass returns, the line below them is how sure it was at every step so far.

from playjev.model import PlayJevModel      # github.com/OmniJev/PlayJev

model = PlayJevModel("OmniJev/PlayJev-0.8B").load()
d = model.decide([frame], options)[0]       # one forward pass, a probability per move
d.choice, d.probs, d.confidence             # the move it takes, the distribution, how sure it is

The prompt, the readout and the option format are built in playjev/model.py, and the demo prints the exact prompt for every game.

🏋️ Training

🤖 Ten teachers one program per game plays on the internal state and labels every frame with a soft target

📸 863k frames behaviour cloning, one epoch, full fine-tuning of the 0.8B base

🔁 Two DAgger rounds the model plays 40k frames per game, the teachers relabel what it visited

One program per game (BFS, expectimax, placement search, A* with deadlock pruning, physics search) plays on the game's internal state and labels frames with a soft target. The model trains on 863k such frames, then on two DAgger rounds where it plays 40k frames per game and the teachers relabel what it visited. Full fine-tuning, one epoch per round, batch 64, learning rate 2e-5, fp32 master weights with bf16 autocast. The trainer is playjev/train_sft.py.

🔭 Open Problems

• 🏃 Motion (Breakout, Mario): one still frame carries no velocity, so the previous frame goes in as a second image.
• 🎯 Single-step precision (Floppy Bird, Tetris): the move that decides the episode is a few hundred frames out of 40k.
• 🔍 Tile digits at 448 px (2048): relabelling cannot help where the digits are unreadable, resolution can.
• ⏱️ Latency: at 83 to 100 ms per step the decision lands one step late, and training on labels shifted one step buys part of it back.

📄 Licence

Apache-2.0, same as the base model. The ten games are other people's work and stay in the code repository under their own licences.