NanoJev Data

The two-stage training inputs and navigation environments for NanoJev, a nano replica of Jev.

Each stage contains all.jsonl and the five split files. all.jsonl preserves the complete training input byte for byte. Split files preserve each original row and its order within the split. Stage 2 continues from the Stage 1 checkpoint and includes replay examples; the two stages are not independent datasets.

Each JSONL row contains id, state_id, family_id, split, a textual state, and a questions mapping. Questions have a type (boolean, choice, or score), instructions, and applicable candidate definitions. Existing per-question target maps, their type annotations, rounding fields, and grouping metadata retain the runtime schema. Candidate order and target values are unchanged.

benchmark/ contains environment-only train, test, and OOD records for the frozen navigation evaluation. Its test and OOD files reproduce the same 40-map selection used in the published comparison. manifest.json records source and output SHA-256 values, exact counts, training settings, and package checks.

Load

Use the repository's JSONL format directly. Rows contain dynamic nested schemas, so a fixed table schema is unnecessary.

import json
from pathlib import Path
from huggingface_hub import snapshot_download

folder = Path(snapshot_download(
    repo_id="C-Tianyu/NanoJev-Data",
    repo_type="dataset",
))

with (folder / "stage2" / "train.jsonl").open() as stream:
    examples = [json.loads(line) for line in stream if line.strip()]

print(len(examples), examples[0]["questions"].keys())

Reading local files needs only Python 3. Downloading needs huggingface_hub and access to this repository. Training uses the NanoJev code, scripts/train_pipeline_decisions.py, and the matching model files. Use stage1/all.jsonl or stage2/all.jsonl as the corresponding --input; the model repository provides the checkpoints. The recorded runtime versions and stage settings are in the manifest.

python verify_dataset.py

This checks all packaged file hashes, split membership and ordering, and the unchanged row content between full inputs and split files.

Game and local-decision inputs

The additive games_v4/ package contains exact inputs for six game and event model variants: local maze questions, game-policy records, frozen target distributions, and probability-event examples. It also includes the complete eight-case Snake cohort and all six greedy/sample runs, with 48 episodes and 9,041 recorded transitions. Compressed rollouts recover the original JSON bytes exactly.

The package preserves all target values and source schemas. Its manifest maps each recorded training configuration to its input hashes. Earlier stage1/, stage2/, and benchmark/ files remain unchanged. See the package README for counts, paths, task definitions, and the role of the shared game planner.

python verify_dataset.py
python games_v4/verify_dataset.py