SolidSnake123/nanochat-world-state-v2-49k-20260714
Nanochat World-State Tracking Capability Each deterministic latent world describes initial people, rooms, portable objects, containers, and fixed surfaces followed by a valid chronological event sequence. The task asks for one exact final location, holder, container, or support. Six natural renderer styles appear in training; validation and test may use all eight. The exact state replay engine supplies every answer. No language model generated or labeled the data. Targeted-v2… See the full description on the dataset page: https://huggingface.co/datasets/SolidSnake123/nanochat-world-state-v2-49k-20260714.
Nanochat World-State Tracking Capability
Each deterministic latent world describes initial people, rooms, portable objects, containers, and fixed surfaces followed by a valid chronological event sequence. The task asks for one exact final location, holder, container, or support. Six natural renderer styles appear in training; validation and test may use all eight. The exact state replay engine supplies every answer. No language model generated or labeled the data. Targeted-v2 training alternates surface-matched invariant and answer-changing suffixes, ordinary base worlds, and timestamp-shuffled reports. Scenario columns are audit metadata and never appear in the training text.
- Train worlds: 49,152
- Validation worlds: 1,024
- Test worlds: 1,024
- Event range: 7-12
- Scenario mode:
targeted_v2 - Dataset format:
nanochat-world-state-dataset-v2 - Latent schema:
world-state-v1 - Renderer:
world-state-natural-v1 - Source commit:
458cf84a6d2a24ba3c26542e024c5b766ca2fea9 - Generation run:
nanochat-world-state-v2-49k-20260714-r1
Only text is consumed by ordinary pretraining. The identity, renderer, query-type, and difficulty columns are audit metadata and are not appended to it.