C-Tianyu/NanoJev-Data

NanoJev-Data — Unified game supervision and recorded evaluation

The complete data package for the current NanoJev model: Maze, Snake, ViZDoom Basic and Predict Position. It includes the exact mixed supervised-learning inputs, full expert episodes, frozen evaluation cohorts, recorded comparisons, and the six-source hard Maze/Snake demonstration.

Training data

The hard and soft variants share the same states, questions and candidates. Only the new Predict Position target changes: expert argmax action versus its complete action distribution. The selected checkpoint uses hard.

Both variants preserve all 7,587 existing Maze, Snake and Basic rows byte for byte. Predict Position contributes 11,173 questions across the five splits, including 6,788 training questions. The raw 896 expert episodes contain 17,498 decisions; the policy-training view retains pre-action states with ammunition, including unsuccessful episodes and the actual firing decisions. Full episodes also retain later states for audit and evaluation.

Rows contain id, state, questions, split, metadata and recorded target annotations. The policy input is only state/question/candidate content; target annotations, terminal outcomes and expert state are excluded from inference. The existing target-validity filter yields 10,893 eligible training questions and quarantines 11 questions over all splits. Stored source rows remain intact.

Expert policy source: Sonic Doom, ordinary no-sound Predict Position visual policy with a recurrent GRU. Basic supervision uses the existing APPO expert collection. Maze and Snake preserve their existing action annotations. Complete provenance stays in each exact source manifest and record; no target annotations are rewritten during packaging.

Evaluation and demonstrations

The frozen benchmark has 548 cases: 274 test and 274 OOD. Four primary systems have complete recordings: current NanoJev, its initialization, Jev and untuned Qwen. All 2,192 episodes pass independent simulator replay. Source manifests, model identities, actual action probabilities and physical counters are included.

Hard demonstrations are separate fixed examples: 50x50 Maze, seed 24310922; 12x12 Snake, seed 61005. They preserve six real model trajectories and the exact common controller definitions. They are not additional samples in the 548-case benchmark. The included current visualization indices are derived artifacts; original recorded trajectories are also present.

See the exact training recipe, dataset counts, and all file hashes. Download the public data package:

import json
from huggingface_hub import hf_hub_download, snapshot_download
repo = "C-Tianyu/NanoJev-Data"
version = "unified-games-v1"
manifest_path = hf_hub_download(repo, "SHA256_MANIFEST.json", repo_type="dataset",
                                revision=version, token=False)
with open(manifest_path) as handle:
    names = [row["path"] for row in json.load(handle)["files"]]
path = snapshot_download(repo, repo_type="dataset", revision=version, token=False,
                         allow_patterns=names + ["SHA256_MANIFEST.json"])

Use the JSONL split files directly with the bundled project training scripts. Split names and bytes match the original experiment. Test/OOD are held out from training and checkpoint selection. Scenario/seed groups are split-disjoint; individual visible observations can repeat between different episodes.

Release history

SHA256_MANIFEST.json describes this complete unified release. The retained root manifest.json and verify_dataset.py describe the earlier dataset layout. Its original snapshot remains available at `legacy-before-unified-games-v1`.