Commit History
-
docs: one row per environment instead of one per tier
Verified signatureObservation and action are the same across an environment's three tiers, so sixteen of the twenty-four rows repeated a spec. Transitions was near constant -- 1,000,000 to 1,012,457 -- which is a sentence, not a column. Episode counts are the part that varies and they now sit side by side, where the direction shows: goal games yield fewer episodes as the tier degrades, locomotion more. Every figure is kept.293a919 -
docs: put the download size next to the download
Verified signatureLoading invited dropping allow_patterns 'to fetch all' without saying that all is 25 GB, or that one environment spans 0.6 to 6.5 GB depending on the scene. The numbers now sit where the choice is made.98b00aa
-
datasets: drop local filesystem paths from collection_spec provenance
Verified signature49da201
-
card: link the collection recipe (examples/unity_collection) in Companion repos
Verified signature972e326
-
Link the GitHub collection/ packager from the download path
Verified signature04a3444
-
Correct companion policy-repo environments to DungeonEscape and SoccerTwos
Verified signature2b1ecbb -
docs: quantify run-to-run precision; state the ladder count plainly
Verified signatureThe intro contrasted "every environment" with "eight ship a ladder" when there are exactly eight environments, so the contrast read as though more existed. The Korean card already said all eight; the English one now matches. The precision paragraph quoted only within-run sampling error. The 2026-07-30 label repair re-recorded the continuous scenes from identical bands and seed, which measures run-to-run spread directly: 0.01-0.03 across the six degraded tiers, with Walker medium and 3DBallHard medium each moving 0.03 - three times the within-run figure quoted for 3DBallHard. Both editions carry it.f2d97b2
-
docs: link the two language editions and correct the checkpoint list
Verified signatureREADME.ko.md was reachable only by browsing the file list -- Hugging Face renders README.md as the card and nothing pointed at the Korean edition, so the full translation had almost no way of being found. Each card now links to the other. Also corrects the companion-repo line from "policy checkpoint (Pyramids)" to "policy checkpoints (Pyramids, PushBlock)": the model repo ships both. The Korean card already said both, so the two editions disagreed until now.0c43435 -
docs(ko): bring the Korean card to full parity with the English one
Verified signatureThe Korean card was about 15% of the English one and, most importantly, carried no data-quality notice at all: a Korean-only reader had no way to learn that 21 of the 24 datasets had been re-recorded, so anyone who downloaded before 2026-07-30 could not know their copy was affected. - Ports the action-label notice in full - Contents table goes from 8 expert-only rows to all 24 datasets - Adds the quality-ladder methodology, the per-environment ladder tables, the measurement-precision note and the episode-length section, all of which the Korean card previously deferred to the English one - Adds the provenance paragraph so the Apache-2.0 (builds/policies) vs CC-BY-4.0 (trajectories) split is stated, not just the latter - Frontmatter aligned: viewer: false and the offline-reinforcement-learning tag Every figure was checked programmatically against the English card rather than retyped.b4503a4 -
docs: document the episode-length distribution of every tier
Verified signatureThe card gave mean episode length per tier but no dispersion, which is the part that actually drives context-window, padding and bucketing decisions for sequence models. Adds a section with episodes / mean / sd / cv / min / median / max for all 24 datasets, measured directly from the published files (local copies verified byte-identical to the shipped LFS objects by sha256). Notes the two regimes and that they move in opposite directions under degradation, the truncation spike that makes mean +/- sd misleading for the strong locomotion tiers, and that DungeonEscape is the only environment whose episodes exceed 1000 steps.e292dbf -
docs: carry the re-recorded continuous figures and close the action-label notice
Verified signatureCrawler, Walker and 3DBallHard were re-recorded on 2026-08-01 with the decision-gap action-label fix, so their episode counts, expert anchors, mean returns, termination rates and tier normalizations all move. Bands and noise seed are unchanged. - Contents / ladder tables / reproduction bullets updated to the measured values - Notice rewritten: all 21 affected datasets are now re-recorded; worm's exemption reason corrected to 'agents never terminate'; the twelve discrete tiers condensed to one row with the 2026-07-30 patch date - Tier spans widened to the measured simple 0.58-0.62 / medium 0.77-0.85, with a note on how precisely those means are known - README.ko.md contents table refreshed (its four discrete rows had gone stale at the 2026-07-30 re-record)1117dc3 -
fix(3dball-hard): re-record all three tiers with the decision-gap action-label fix
Verified signatureRe-recorded with the 99a8b30 collector, which carries the last decision forward across ML-Agents decision gaps instead of storing a placeholder 0. Bands and noise seed are unchanged; episode counts and tier metrics shift because a collection run is not bit-reproducible.741f4b4 -
fix(walker): re-record all three tiers with the decision-gap action-label fix
Verified signatureRe-recorded with the 99a8b30 collector, which carries the last decision forward across ML-Agents decision gaps instead of storing a placeholder 0. Bands and noise seed are unchanged; episode counts and tier metrics shift because a collection run is not bit-reproducible.cda5259 -
fix(crawler): re-record all three tiers with the decision-gap action-label fix
Verified signatureRe-recorded with the 99a8b30 collector, which carries the last decision forward across ML-Agents decision gaps instead of storing a placeholder 0. Bands and noise seed are unchanged; episode counts and tier metrics shift because a collection run is not bit-reproducible.dfcd86e -
card: state the continuous scenes as published, with measured placeholder shares
Verified signatureThe continuous scenes were never scheduled for a re-record; the notice listed them as pending anyway. Replaces that status with a factual disclosure and swaps the estimated 1% for shares counted directly in the published files: crawler 0.57-1.03%, walker 1.05-1.62%, 3dball-hard 0.02-1.46%, worm 0% across all three tiers. Discrete figures are unchanged and remain replay estimates of the files that were replaced.622aa4a -
docs: add Crawler to the reproduction recipe
Verified signatureThe Crawler build and its stock ONNX are published in the envs repo again, so Crawler follows the same stock-ONNX recipe as the other continuous scenes. Expert validation figures measured from the shipped dataset: 1,048 episodes, 1,000,002 transitions, mean return 2576.4899, mean length 954.20, 81 terminated / 967 truncated.9362a62 -
card: correct the realized tier-norm ranges and the pending-scene count
Verified signatureThe overview quoted simple 0.59-0.61 / medium 0.80-0.82; the shipped tables are simple 0.60-0.61 and medium 0.79-0.82 (pyramids). The data-quality notice also called three pending continuous scenes four while excluding worm in the same sentence.69f3b4d -
docs: sync soccer-twos and dungeon-escape tier figures to the re-recorded datasets
Verified signaturereproduction.md was last updated after the expert re-records but before the medium/simple ones, so its tier counts and dungeon-escape norms predated the shipped files and contradicted the card.9d0777c
-
card: all 12 discrete-env datasets re-recorded for the action-label gap; refresh counts and ladder figures
Verified signaturee8048de -
pyramids/simple-v0: re-record with the action-label gap fix (2.98% of rows held a placeholder 0); band U(4.8,7.1) unchanged
Verified signatureac3dd77 -
pyramids/medium-v0: re-record with the action-label gap fix (4.46% of rows held a placeholder 0); band U(2.7,4.9) unchanged
Verified signature52866d2 -
pushblock/simple-v0: re-record with the action-label gap fix (46.6% of rows held a placeholder 0); band U(3.25,4.75) unchanged
Verified signature08db4f4 -
pushblock/medium-v0: re-record with the action-label gap fix (61.0% of rows held a placeholder 0); band U(2.1,3.2) unchanged
Verified signature23025f5 -
dungeon-escape/simple-v0: restore the band-qualified policy_id (stock-r23-group-temperature-range-3.0-4.0) dropped in the action-label re-record
Verified signatureb13636e -
dungeon-escape/medium-v0: restore the band-qualified policy_id (stock-r23-group-temperature-range-1.5-3.0) dropped in the action-label re-record
Verified signaturedd28397 -
soccer-twos/simple-v0: re-record with the action-label gap fix (6.4% of rows were a placeholder [0,0,0]); per-team band U(2.1,3.4) and side-swap norm ~0.60 unchanged
Verified signature755c836 -
soccer-twos/medium-v0: re-record with the action-label gap fix (6.4% of rows were a placeholder [0,0,0]); per-team band U(1.3,2.6) and side-swap norm ~0.80 unchanged
Verified signature0d82029 -
dungeon-escape/simple-v0: re-record with the action-label gap fix (action 0 share 32.49% -> 11.05%); band U(3.0,4.0) and norm 0.5990 unchanged
Verified signatureacfbabc -
dungeon-escape/medium-v0: re-record with the action-label gap fix (action 0 share 36.93% -> 9.53%); band U(1.5,3.0) and norm 0.8007 unchanged
Verified signature348b284 -
docs: pyramids + soccer-twos expert validation numbers after re-record
Verified signaturee8c07bd -
card: pyramids + soccer-twos expert-v0 re-record (all four discrete experts now fixed)
Verified signaturee7ac486 -
soccer-twos/expert-v0: fix action labels — the collector recorded a fabricated (0,0,0) for every step an agent spent inside ML-Agents' action-repeat gap (11.19% of rows held (0,0,0) where the policy's joint probability for it is 3.80%); it now carries the last real decision forward, so the action column matches the policy the trajectories were actually driven by
Verified signaturee916ceb -
pyramids/expert-v0: fix action labels — the collector recorded a fabricated 0 for every step an agent spent inside ML-Agents' action-repeat gap (~6% of rows); it now carries the last real decision forward, so the action column matches the policy the trajectories were actually driven by
Verified signature228e2a0 -
docs: dungeon-escape expert validation numbers after re-record
Verified signature7548bc0 -
card: dungeon-escape expert-v0 re-record (35,505 ep, anchor 0.019355, medium norm 0.81)
Verified signature40c9b73 -
dungeon-escape/expert-v0: fix action labels — the collector recorded a fabricated 0 for every step an agent spent inside ML-Agents' action-repeat gap (41.5% of rows); it now carries the last real decision forward, so the action column matches the policy the trajectories were actually driven by
Verified signatureb5d1710 -
docs: pushblock expert re-record numbers + decision-period action-label rule
Verified signature6a5ca27 -
card: pushblock expert-v0 re-record (53,039 ep, anchor 0.353699) + action-label data-quality notice
Verified signature564e6c2 -
pushblock/expert-v0: fix action labels — the collector recorded a fabricated 0 for every step an agent spent inside ML-Agents' action-repeat gap (72.16% of rows); it now carries the last real decision forward, so the action column matches the policy the trajectories were actually driven by
Verified signature364d655 -
docs(card): correct the companion policy-checkpoint list to Pyramids
Verified signatureThe model repo ships one Causal GPT-RL policy, pyramids/pyramids-b16.onnx. The card still listed Crawler, PushBlock, SoccerTwos and DungeonEscape checkpoints; Crawler was removed and the other three were never published there.a1a06ac
-
docs: frame as offline-RL datasets (usable by any offline-RL method); add offline-reinforcement-learning tag
Verified signature8f63f6a -
docs: add 'reduced backward from the expert' rationale to Quality ladders intro
Verified signatured13cf4c -
fix: replace attached '~' (approx) with ≈ so HF markdown stops rendering it as strikethrough; backtick 'T ~ U' notation
Verified signature2ec3ee6 -
fix: replace attached '~' (approx) with ≈ so HF markdown stops rendering it as strikethrough; backtick 'T ~ U' notation
Verified signaturea873bcd -
fix: replace attached '~' (approx) with ≈ so HF markdown stops rendering it as strikethrough; backtick 'T ~ U' notation
Verified signatureaf9985d -
fix: pyramids medium/simple episode counts in dataset table (4267->4159, 3040->3042) to match re-collected tiers
Verified signaturee1b5fc4 -
docs: intro describes softmax-temperature degradation for discrete tiers (not just action noise); refresh simple/medium spans
Verified signature5175018 -
soccer-twos medium/simple: describe per-team per-match softmax-temperature-range self-play tiers (T~U(1.3,2.6)/U(2.1,3.4), side-swap norm ~0.80/0.60), replacing the epsilon-pool description
Verified signature07088c6 -
soccer-twos medium/simple: describe per-team per-match softmax-temperature-range self-play tiers (T~U(1.3,2.6)/U(2.1,3.4), side-swap norm ~0.80/0.60), replacing the epsilon-pool description
Verified signaturef2d08b3 -
soccer-twos medium/simple: describe per-team per-match softmax-temperature-range self-play tiers (T~U(1.3,2.6)/U(2.1,3.4), side-swap norm ~0.80/0.60), replacing the epsilon-pool description
Verified signatureb9dcb9b