novastar111/uwm_paper_artifact_inventory

UWM Hugging Face artifact inventory

Audit date: 2026-08-24 UTC. Primary public organization: novastar111; quota-overflow checkpoint weights: JiaxinGe/umw.

Public index: novastar111/uwm_paper_artifact_inventory.

Paper evaluation outputs: uwm_paper_eval_trajectories. Its README maps every planned Table 1/2/3 result cell to a full-run trajectory archive. Currently The original release index is 230/230 complete across the two release repositories, including all 24/24 canonical Vanilla-3k → 2-box RL trajectories and all 24/24 Vanilla-3k → 3-box RL trajectories at steps 50/100/150/200. The final Pacman release contains 44 completed 100-episode trajectory archives, including both Easy random-100 v1 and Easy stratified-100 v2 and the Forward-v2 g8/g12 stress runs. CoT/forward/branch archives contain the model's generated observation images; action-only NonCoT protocols have n_imagined=0 by construction and contain the complete actions, environment renders, GIFs, and episode statistics instead.

Kinf NonCoT is the single-shot/open-loop NonCoT arm for Pacman and PushT. Every hyperlink below except the rows explicitly marked pending artifact was checked against the public Hub inventory. The base checkpoint/data manifest passed 157/157. The adaptive RL release now includes the complete 10-step checkpoint ladder from RL10 through RL200 for Vanilla-5k, Mixed-Forward-3k, and MULTIB3-3k.

Table 1

Checkpoints

Task	K1 NonCoT	K3 CoT	K3 NonCoT	K5 CoT	K5 NonCoT	K10 CoT	K10 NonCoT	Kinf CoT	Kinf NonCoT
Sokoban	model	model	model	model	model	model	model	model	model
Pacman (New braided/thinned)	model	model	model	model	model	model	model	model	model
Maze2D	model	model	model	model	model	model	model	model	model
PushT (New diverse-start)	model	model	model	model	model	model	model	model	model

PushT CoT K3/K5/K10 use the final diverse_keeplast checkpoints. CoT Kinf uses the diverse checkpoint whose training configuration has vae_last_only_attn=True. PushT K1 is NonCoT only; the Kinf NonCoT column is the single-shot/open-loop checkpoint.

Training data

Task	K1 NonCoT	K3 CoT	K3 NonCoT	K5 CoT	K5 NonCoT	K10 CoT	K10 NonCoT	Kinf CoT	Kinf NonCoT
Sokoban	data	data	data	data	data	data	data	data	data
Pacman (New braided/thinned)	data	data	data	data	data	data	data	data	data
Maze2D	data	data	data	data	data	data	data	data	data
PushT (New diverse-start)	data	data	data	data	data	data	data	data	data

Evaluation data

Task	Eval set	Link
Sokoban	Ordinary easy, 100 episodes	sokoban_easy_eval_100_20260712
Sokoban	Ordinary hard, 4 boxes	sokoban_hard_init_state_pool
Pacman (New)	Easy, deterministic random 100 v1	pacman_eval_easy_braidthin_random100_v1_20260820
Pacman (New)	Easy, distribution-stratified 100 v2	pacman_eval_easy_braidthin_stratified100_v2_20260820
Pacman (New)	Ordinary hard, deterministic random 100	pacman_eval_hard_braidthin_random100_20260817
Pacman (New)	Two-Ghost, deterministic random 100	pacman_eval_two_ghost_braidthin_random100_20260817
Maze2D	Ordinary hard, larger maze	plain / CoT
Maze2D	Ordinary easy	plain / CoT
PushT (New)	In-dist, balanced 100; exact-state disjoint from training	pusht_eval_in_dist_balanced100_20260820
PushT (New)	Near diagnostic, balanced 100	pusht_eval_near_balanced100_20260810
PushT (New)	Mid diagnostic, balanced 100	pusht_eval_mid_balanced100_20260810
PushT (New)	Far strict spatial OOD, balanced 100	pusht_eval_far_balanced100_20260810

Table 2

Checkpoints

Task	Variant	Link
Sokoban	Vanilla, exact SFT step 3000	sokoban_adaptive_vanilla_sft3k
Sokoban	Always Forward Thinking, step 5000	sokoban_easy_cot_chunk_kinf_compare2
Sokoban	Ordinary Branching, step 5000	sokoban_easy_cot_chunk_kinf_branch
Sokoban	Mixed Forward Thinking, exact step 3000	sokoban_adaptive_mixed_forward_sft3k
Sokoban	Mixed Forward Thinking, exact step 5000	sokoban_adaptive_mixed_forward_sft5k
Sokoban	Mixed Branching, exact step 3000	sokoban_adaptive_mixed_branch_sft3k
Sokoban	Mixed Branching, exact step 5000	sokoban_adaptive_mixed_branch_sft5k
Pacman	Always Forward Thinking	pacman_braidthin_always_forward_kinf_world_model
Pacman	Forward Thinking v2, safe-longer comparisons	checkpoint
Pacman	Forward Thinking v3, mixed fatal/safe comparisons	checkpoint
Pacman	Forward Thinking v4, short-trap-only comparisons	checkpoint

Training data

Task	Variant	Link
Sokoban	Always Forward Thinking	sokoban_adaptive_always_forward_train
Sokoban	Ordinary Branching	sokoban_adaptive_ordinary_branch_train
Sokoban	Mixed Forward	sokoban_easy_cot_kinf_forward_random_train
Sokoban	Mixed Branching	sokoban_easy_cot_kinf_branch_multi_train
Pacman	Always Forward Thinking	pacman_braidthin_always_forward_kinf_train
Pacman	Forward Thinking v2, cutoff 50	data
Pacman	Forward Thinking v3, mixed fatal/safe	data
Pacman	Forward Thinking v4, short-trap-only	data

Evaluation data

Task	Eval set	Link
Sokoban	2-box Deadlock	dataset
Sokoban	3-box Deadlock	dataset
Sokoban	4-box Deadlock	dataset
Sokoban	2-box Branch / Structural Fork	dataset
Sokoban	3-box Branch / Structural Fork	dataset
Sokoban	4-box Branch / Structural Fork	dataset
Pacman	Ordinary hard, deterministic random 100	dataset
Pacman	Two-Ghost, deterministic random 100	dataset
Pacman	Forward stress G8, deterministic random 100	dataset
Pacman	Forward stress G12, deterministic random 100	dataset

Pacman route-order diagnostics and trajectories

The Pacman route-order stress test is built around a certified delayed-consequence fork rather than an immediately dangerous action. Each episode starts with a food-neutral shared corridor leading to the same decision state. The attractive branch reaches a statically nearer food and initially survives, but its continuation toward the remaining objectives is verified to collide with the deterministic ghost and bounded search certifies that it cannot finish within the episode budget. The alternative branch first targets a farther food, yet has a replay-verified global plan that collects everything alive and executes STOP. G8/F2, F4, F6, and F8 isolate increasing objective count; G10/F8 additionally increases spatial extent as a combined out-of-distribution stress test. Full visual artifacts include generated imagined frames, environment renders, rollout GIFs, generated text/actions, and episode statistics.

Evaluation	Vanilla	Forward-v3	Forward-v4	Trajectories / test data
Original G8/F2 route-order set	2/100	9/100	2/100	full visual runs
Original G8/F4 route-order set	7/100	11/100	7/100	full visual runs
Original G8/F6 route-order set	7/100	10/100	9/100	full visual runs
Strict topology-first G8/F2 v2	7/100	12/100	40/100	full run + visual comparisons
Strict topology-first G8/F4 v2	22/100	15/100	54/100	full runs + exact test set
Strict topology-first G8/F6 v2, prefix 2--9	16/100	20/100	41/100	full runs + exact test set
Strict topology-first G8/F8 v2, prefix 2--8	16/100	15/100	37/100	full runs + exact test set
Strict topology-first G10/F8 v2	5/100	5/100	12/100	full runs + exact test set

The controlled extension keeps the grid at G8 and varies only food count. F6 omits shared-prefix length 10 because two independent searches of one million candidates found no valid example under the strict trap constraints; lengths 2--9 remain represented. G8/F8 uses prefix lengths 2--8 after prefix 9 yielded zero examples in roughly 100,000 candidates. The G8 suites have zero exact initial-state and wall-layout overlap. The v4 training set (G9/F8) also has zero exact or dihedral-normalized state/layout overlap with strict F2; this rules out sample-level leakage, while the task concept itself is intentionally aligned with the short-trap training objective. For G10/F8, food-complete/alive scores that ignore a missed terminal STOP are 6/100, 5/100, and 15/100 respectively; the table consistently retains the stricter stop-required metric.

The original selected trajectories are also available in the standalone novastar111 qualitative release. The earlier JiaxinGe release paths remain available for route-order v1 and the strict F2 v2 paired runs.

The six Sokoban links above point to the exact sokoban_paper_final_v2_max40 shards used by the published trajectory runs: 100 episodes per set, oracle/expert solution length at most 20, model execution budget 40, q95/perseg full self-rollout, move-only actions, and stop-required success. Their remote LFS hashes were verified exactly against the local frozen suite (6/6 PASS).

The already-published actual Sokoban model-output trajectories for the earlier four-set comparison are at adaptive_thinking_final_v1.

Table 3 — RL

Checkpoints

The paper-facing RL comparison uses Vanilla SFT-5k, Mixed-Forward SFT-3k, and MULTIB3 SFT-3k. Every saved 10-step checkpoint through RL200 is published:

RL step	Vanilla-5k	Mixed-Forward-3k	MULTIB3-3k
10	model	model	model
20	model	model	model
30	model	model	model
40	model	model	model
50	model	model	model
60	model	model	model
70	model	model	model
80	model	model	model
90	model	model	model
100	model	model	model
110	model	model	model
120	model	model	model
130	model	model	model
140	model	model	model
150	model	model	model
160	model	model	model
170	model	model	model
180	model	model	model
190	model	model	model
200	model	model	model

Complete all-arms RL50/RL100/RL150/RL200 trajectories

JiaxinGe/umw/sokoban_rl_allarms contains 72/72 full self-rollout archives: three arms × four RL steps × six frozen evaluation sets. Each archive contains 100 episode records, 100 GIFs, environment renders, raw generated text/actions, and model-generated gen_obs/imagined*.png. All episodes are validated as imagine_mode=self_rollout. This expanded release adds 54 distinct RL50/RL100/RL150 run cells to the original 230-run inventory and mirrors the 18 RL200 cells, giving 284 unique evaluated run cells overall.

Arm	RL step	2D Deadlock	3D Deadlock	4D Deadlock	2-Box Branch	3-Box Branch	4-Box Branch
Vanilla-5k	50	20	41	33	84	37	24
Vanilla-5k	100	22	42	34	85	43	30
Vanilla-5k	150	22	43	34	86	45	31
Vanilla-5k	200	22	44	38	87	44	35
Mixed-Forward-3k	50	35	42	27	75	44	27
Mixed-Forward-3k	100	21	43	33	69	46	28
Mixed-Forward-3k	150	24	48	35	76	43	28
Mixed-Forward-3k	200	24	45	40	77	47	32
MULTIB3-3k	50	36	37	30	84	36	25
MULTIB3-3k	100	32	44	40	85	44	34
MULTIB3-3k	150	34	48	49	84	46	40
MULTIB3-3k	200	34	53	53	84	50	45

RL training reward curves and raw metrics

The complete unsmoothed training metrics for the three primary RL arms are published at JiaxinGe/umw/training_metrics/sokoban_rl_allarms. All three runs cover every training step from 1 through 200.

Artifact	Coverage	Link
Plot-ready reward curves	600 rows: 3 arms × 200 steps; unsmoothed	reward_curve.csv
Rank-level rollout rewards	9,600 rows: 3 arms × 200 steps × 16 ranks	rank_rollout_rewards.csv
Full trainer metrics	9,600 process-level rows	step_metrics_all_processes.csv
Raw reward log records	19,200 original log lines	raw_reward_log_lines.jsonl.gz
Integrity metadata	Exact byte sizes and SHA-256 digests	manifest.json

The trainer did not retain per-episode reward identities. The finest preserved reward granularity is therefore one row per rank and training step, summarizing eight rollouts; this is the content of rank_rollout_rewards.csv.

The canonical Vanilla SFT-3k checkpoint is the intermediate step from the same deterministic run that continues to SFT-5k. Its matched 2-box and 3-box RL runs are complete through RL200. To avoid the ambiguous model versus weight labels used by an earlier revision, every entry below is called a checkpoint, regardless of whether it is hosted in a model repository or a dataset subdirectory. The dense ladders are hosted under JiaxinGe/umw because the original namespace reached its public LFS quota.

RL step	Vanilla-3k → 2-box RL	Vanilla-3k → 3-box RL
10	checkpoint	checkpoint
20	checkpoint	checkpoint
30	checkpoint	checkpoint
40	checkpoint	checkpoint
50	checkpoint	checkpoint
60	checkpoint	checkpoint
70	checkpoint	checkpoint
80	checkpoint	checkpoint
90	checkpoint	checkpoint
100	checkpoint	checkpoint
110	checkpoint	checkpoint
120	checkpoint	checkpoint
130	checkpoint	checkpoint
140	checkpoint	checkpoint
150	checkpoint	checkpoint
160	source checkpoint unavailable (pruned by the training run's top-k retention policy)	checkpoint
170	checkpoint	checkpoint
180	checkpoint	checkpoint
190	checkpoint	checkpoint
200	checkpoint	checkpoint

RL evaluation trajectories (with generated images)

The complete per-episode archives are published in two locations:

• RL2 RL50/RL100/RL150/RL200 and RL3 RL50/RL100/RL150
• RL3 RL200 — all six sets

Each archive includes the action trace, environment renders, generated observation images, episode_stats.json, and rollout.gif. Together these cover RL50/RL100/RL150/RL200 on the six frozen 2/3/4-box deadlock and structural-fork sets (100 episodes per cell), 48/48 archives.

Training data

Data	Link
3-box RL set	sokoban_adaptive_rl_3box_train_2000
3-box SFT control set	sokoban_compute_matched_3box_sft_train_2000
2-box RL set	sokoban_adaptive_rl_2box_train_2000

Evaluation data

Table 3 uses the same six frozen Sokoban datasets linked in Table 2.

Original-data continued-SFT compute control

Initialization: canonical 5k-horizon SFT checkpoint at step 3000. Training resumes the exact deterministic cursor of the original 100,000-example ordinary Vanilla Kinf dataset. The 17,096 continuation states have zero overlap with the 24,008 states consumed through canonical step 3000, and the dataset does not wrap.

SFT updates	RL compute match	Checkpoint	2D	3D	4D	2-Branch	3-Branch	4-Branch	Trajectories
534	50	checkpoint	28	33	31	86	34	15	full images
1068	100	checkpoint	27	32	35	85	33	18	full images
1602	150	checkpoint	29	32	33	85	33	19	full images
2136	200	checkpoint	28	33	34	85	34	19	full images

Training data: https://huggingface.co/datasets/novastar111/sokoban_easy_cot_chunk_kinf_train