The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
SPAgent Non-FrozenLake 2K
Training data and environment code for five deterministic, fully observed multi-turn tasks. FrozenLake is intentionally excluded.
| Environment | Train rows | Unique states | Repeated rows |
|---|---|---|---|
VRBench/Pathfinder-v0 |
2,000 | 2,000 | 0 |
VRBench/Maze-v0 |
2,000 | 2,000 | 0 |
VRBench/TrapField-v0 |
2,000 | 2,000 | 0 |
Games/Sokoban-v0 |
2,000 | 1,600 | 400 |
Games/FifteenPuzzle-v0 |
2,000 | 2,000 | 0 |
TrapField was generated by scanning 1,301,248 official Beyond-MMEnv seeds with 16 workers and retaining only unique solvable states.
Interaction Contract
All tasks use the model-facing SPAgent/FrozenLake-v3 contract:
- explicit initial image placement after
# Images; <think>followed by<tool_call>or<answer>;move(direction=Left|Down|Right|Up);- cumulative image anchors in tool responses;
- complete conversation history on every rollout turn;
- final action-sequence replay from the initial state for reward.
Task rules, grid symbols, and transition mechanics remain task-specific.
Layout
data/
vrbench_navigation_v1/
*_train_v1.jsonl
images/
grid_puzzles_v1/
*_train_v1.jsonl
images/
code/
spagent/
train/
test/
Image paths inside the uploaded JSONL files are relative to each JSONL directory, so the dataset does not depend on paths from the generation machine.
Training
Set MODEL to the local SFT checkpoint before launching:
MODEL=/path/to/sft-checkpoint \
DATA_ROOT="$PWD/data/vrbench_navigation_v1" \
TASK=mixed \
bash code/train/train_vrbench_navigation_grpo.sh
MODEL=/path/to/sft-checkpoint \
DATA_ROOT="$PWD/data/grid_puzzles_v1" \
TASK=mixed \
bash code/train/train_grid_puzzles_grpo.sh
The VRBench mixed file contains 6,000 records. The grid-puzzle mixed file contains 4,000 records. Evaluation is maintained separately as a 30-episode per-environment suite.
- Downloads last month
- 43