Zeb-42 (February 2026): self-play example batches
The batches of self-play examples that Vast.ai workers uploaded while training the February 2026 Zeb, the
AlphaZero-style Texas 42 player in jasonyandell/mk5-main at forge/zeb/.
Jason Yandell and Claude agents built it between 2026-01-31 and 2026-02-17. It is the sibling of the model repo
jasonyandell/zeb-42 (MIT), and the Hub was the bus between workers and
the learner: workers pulled weights and pushed example files here, and the learner consumed them. The repo holds
3,418 files, 20.6 GB, created 2026-02-07. Public domain (CC0 1.0).
Name warning. "Zeb" here is the February 2026 self-play player, a 557K to 3.3M parameter transformer trained with MCTS. Since 2026-10-10 the project uses "Zeb" for the family of players built on Walt, the scenario player, which has nothing to do with these files. The old meaning is kept as written in the zeb page.
Map
Every .pt file is one worker upload, named <worker>_<unix seconds>_<hash>.pt. File contents are not documented in
the wiki beyond "examples"; the zeb-42 model card calls them examples.pt. The date ranges below come from the
timestamps in the file names. Which model each directory trained is only partly documented, as noted.
| directory | files | dates (UTC, from file names) | what it is | explained by |
|---|---|---|---|---|
| (repo root) | 15 | 2026-02-07 | the earliest worker uploads. Model not documented in the wiki. | zeb, zeb-fleet-ops |
large/ |
43 | 2026-02-08 to 02-09 | worker uploads for the run named large. Not documented in the wiki beyond the name. |
zeb |
large-belief/ |
75 | 2026-02-11 | worker uploads for the run named large-belief, the 3.3M-parameter belief-head model. |
zeb |
zeb-large-belief-eq/ |
21 | 2026-02-12 | files named eval-stage1-seed_*, not worker self-play. Not documented in the wiki. |
not documented in the wiki |
lb-v-eq-3740/ |
160 | 2026-02-16 | examples from the 2026-02-15 run bootstrapped from large-belief.pt at Hub step 3740: one learner, two self-play workers and an eval-aux worker |
zeb-fleet-ops (the worker saga) |
lb-v-eq-1920/ |
3,103 | 2026-02-16 to 02-17 | examples from the later full-teacher run, mixing oracle E[Q] teacher examples with self-play. This is a different run from lb-v-eq-3740. |
full-teacher-eq-experiment |
The oracle era
The teacher in lb-v-eq-1920 was the perfect-information oracle's E[Q], which the project has since fenced off; see
the strategy-fusion fence.
What it is the input of
The learner for the weights in zeb-42. The full-teacher run's outcome, that E[Q] as a teacher does not move play past the self-play ceiling, is on the experiment page above. Nothing newer is trained on these files.
Links
Zeb entity page, Hub inventory, GitHub repo.
License
CC0 1.0 Universal (public domain dedication); the legal code is in LICENSE.