NanoJev — A nano replica of Jev

One 0.6B model returns action distributions for Maze, Snake, ViZDoom Basic and Predict Position. This public release contains the current unified step-400 model, its exact training initialization and the full inference source. Jev inspired the state, question and candidate-set interface.

Each candidate is encoded with its state and question. A shared Qwen3-0.6B backbone, end-of-sequence representations and an attention-based Choice head score the offered set. Softmax returns all candidate probabilities together, without output-token decoding. Boolean and ordered-score request types remain available through the same structured interface.

Matched test task	NanoJev	Jev	Untuned Qwen
Maze	4/10	7/10	2/10
Snake	8/8	8/8	0/8
Basic	128/128	56/128	56/128
Predict Position	27/128	11/128	11/128

All systems share the recorded structured observation interface, frozen cases, controller epsilon 0.1 and sampling seed 17. Full test and OOD reports, source trajectories and independent replay checks are in the linked dataset. The hard 50x50 Maze and 12x12 Snake demonstrations are a separately recorded showcase using common code planners: current NanoJev reaches the maze goal in 225 attempts and collects 30 food items while surviving all 256 Snake steps.

Load the model

Complete mixed dataset and recordings · Training recipe · File hashes

The release is available without authentication. The root directory is the current complete checkpoint bundle; it is loaded by NanoJev's DecisionPredictor.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("C-Tianyu/NanoJev", revision="unified-games-v1",
                         token=False, allow_patterns=["best.safetensors", "config.json",
                         "backbone_config/*", "tokenizer/*", "source/*"])
sys.path.insert(0, path + "/source/scripts")
from predict_toy_decisions import DecisionPredictor

model = DecisionPredictor(path, device_name="cuda:0", precision="bf16",
                          disable_native_triton=True)
request = {"states": [{"id": "example", "state": "The target is left of the aim.",
    "questions": {"action": {"type": "choice", "instructions": "Choose the next action.",
        "criteria": {"left": "Move the aim left.", "right": "Move the aim right.",
                     "shoot": "Fire.", "noop": "Wait."}}}}]}
answer = model.predict(request)
print(answer)

The saved format contains best.safetensors, config.json, backbone_config/ and tokenizer/, plus the original training audits. It is the project's decision model format. The base revision is Qwen/Qwen3-0.6B@c1899de289a04d12100db370d81485cdf75e47ca.

Training uses mixed-task supervised cross entropy. Predict Position action supervision comes from the released Sonic Doom visual expert. The existing Maze, Snake and Basic data remain in their original splits. Four matched runs compare hard/soft PP targets and two learning rates; development performance selects hard_lr1e5 at step 400. The selected weight SHA256 is f68c47d66998231b86b7e91b4ed5e82ae23acf104c8b7cd6d165c3ac7b7ffe1b. The full original 15-file bundle is preserved.

Source code carries the included MIT license; the Qwen base model retains its upstream license. Model and data snapshots are published as unified-games-v1.

Release history

The inference download above includes the current model and source. The full training package also includes training_initialization/; use the manifest allowlist in the recipe to download all current files.

The current release is verified by SHA256_MANIFEST.json. Retained MODEL_MANIFEST.json, GAMES_MODEL_MANIFEST.json, stage1/ and variants/ belong to earlier releases. The prior root snapshot remains available at legacy-before-unified-games-v1.