language:
- en
task_categories:
- image-to-video
COF new-task data snapshot
Private data companion to the ongoing four-task AR recovery experiment. Code: https://github.com/mianwu01/COF-maze/tree/new-tasks-transfer-2026-10-08-14-28-55 . No final two-round success claim is made.
Contents
symbolic_records_and_metadata.tar.gz: exact packing, Sudoku and sliding records (4000 train / 32 dev / 250 test each, plus an 8-question diagnostic subset of dev); Maze-v2 symbolic train/test corpus, independent dev128 and diagnostic8; Maze-v2 R1 correction plan, gates and provenance metadata.maze_v2_base_latents-*.tar: the 1854 fixed VAE-encoded clean training examples and original manifest/encoding metadata.maze_v2_r1_corrections-*.tar: all 473 accepted first-round corrections, preserving raw policy-prefix latents, supervision masks and context-encoded expert suffixes, with their original gates/manifests.manifest.json: SHA256 and sizes of the uploaded archives.
The Maze-v2 base training population has 1854 records; its frozen test population has 1000. Development and diagnostic populations are distinct from that test population. The diagnostic8 sets are early probes, not independent full evaluations.
Why retain latents
Symbolic records can regenerate expert demonstrations. They cannot recreate a specific policy-generated prefix just by rendering the expert solution. The accepted R1 latent files retain those prefixes and their expert continuations, so the fixed correction training dataset can be reused without repeating collection.
The base plus R1 mixture has 1854 + 473 = 2327 examples. Use the original source code to mix these files and build a configuration; do not treat metadata as portable training configs. Some original provenance metadata contains absolute paths. Keep the original metadata for verification, and construct paths and new run identities for the destination machine rather than silently editing a hash-bound manifest.
Verification and unpacking
Verify each archive's SHA256 against manifest.json, then extract all archives into the same empty data root. Paths inside the tar archives identify task_records/, maze_v2/symbolic/, maze_v2/dev128/, maze_v2/diagnostic8/, maze_v2/base_latents/ and maze_v2/r1_corrections/. The compressed JSONL task records require a second gzip decompression for inspection.
These data support reproducing the symbolic corpus and reusing the fixed base/R1 training data. Exact optimizer continuation still requires the appropriate full model/trainer checkpoint and world size 4. GPU kernels are not guaranteed bitwise deterministic.
Exclusions and snapshot boundary
No model checkpoints, environment archives, unrelated task logs or credentials are included. Full raw rollout latents before correction filtering are not included; rejected-trajectory recollection and planning are therefore not fully reconstructible from this package alone. Base latent caches for the other three tasks are not included; they can be rebuilt from the published symbolic records. Other tasks' R1 data and Maze-v2 R2 data are not complete at this snapshot and are not claimed to be present. The archive manifest and code commit identify the fixed release.