diff --git a/latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json b/latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json index 5806c9600ff7d3debdf61b7a05314a76cd5c430d..f65c901627c159bf5cf1293eac8c01635eeed27a 100644 --- a/latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json +++ b/latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json @@ -7,29 +7,29 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/checkpoint_p0/checkpoint_000048836_25004032.pth" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/git.diff" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/sf_log.txt" } ], diff --git a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json index 807534c5888846881fd1ca25fa503be68357e1e2..5d7272b7bcf27e34c46334a51d480d8969e1ee5d 100644 --- a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/summary.jsonl" } ], @@ -72,7 +72,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" } diff --git a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json index 462106ba09f1cf624f048929024cc3bd35580e52..334df2bad7ec36de56687beefe35c9c27c2aa526 100644 --- a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/summary.jsonl" } ], @@ -72,7 +72,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" } diff --git a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json index 9f098aeded14bf7a6ed99e528c9071dd4160da34..7d6e56b5ede37e3702337ff1d0c60449c5e4a92c 100644 --- a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json @@ -7,49 +7,49 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval_inputs/deadly_corridor_fixed_latency_6.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -82,7 +82,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" } diff --git a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json index c66e1db0a6a4033e1ff063bcb593939dbe9b9236..f6f0d066835e1f6488d88ec1cdbc5c12e8f332ec 100644 --- a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/summary.jsonl" } ], @@ -72,7 +72,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" } diff --git a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json index 2984e823051eef0319091d8b6a78eb47abb30ca2..45b651979fc7732074d5f46e37d50401e5fb2f6e 100644 --- a/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/steps_500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "deadly_corridor_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json b/latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json index 684ec9519749736c94e25a4128f56b35c29562f3..fa7c7db990ba67f209b5db82a7de6ec23ef49ddb 100644 --- a/latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json +++ b/latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json @@ -7,29 +7,29 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/eval_latency_6/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/queue_eval_latency_6.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/eval/fixed_6/queue_eval_results.jsonl" } ], @@ -62,7 +62,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" } diff --git a/latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json b/latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json index c7951c17cb28c088d5b7bc842b1a0222a13a813d..8d5a87b0bae7c7de4489673edea5a8e364676192 100644 --- a/latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json +++ b/latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json @@ -7,29 +7,29 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/checkpoint_p0/checkpoint_000048880_25026560.pth" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/git.diff" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/sf_log.txt" } ], diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json index 2ddccc266bc64e67bf2b9714f804c99ca408042c..e7ba47ba729b5581591a95adf7c5d07930be88c5 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/checkpoints/steps_7000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "demon_attack_fix_latency_2_200ep", diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json index 525c1a24cc9b6cfb4bf99bba47af4f4f6ca49290..c720662bd38bc4c70c22867ab5185ce84ae07940 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/checkpoints/steps_7000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "demon_attack_fix_latency_4_200ep", diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json index 411d9276b5caadd3396dfffad9b61b3b44001bf0..366337f2421efa564851c1f1de414dd666fad207 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/checkpoints/steps_7000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "demon_attack_fix_latency_6_200ep", diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json index c5a494eb4af0ffbac9d77ee304a053b34866a94a..2a779694c8468a2dd669f117668ee6ed815aa1eb 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json @@ -7,44 +7,44 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.cap3600.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/summary.jsonl" } ], @@ -77,7 +77,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" } diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json index fe87514c7fd027061cb1a717c109f772927b3681..8c88486018e10857aaca7c9cfa08d28ff45343d1 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json @@ -7,49 +7,49 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.cap3600.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.pre_framealign.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/summary.jsonl" } ], @@ -82,7 +82,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" } diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json index 802b60b4d03715fb21b80a451dbb9ed70003c3c6..cc0d1ff400697986b174f22115884faa9c0a2036 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json @@ -7,44 +7,44 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.cap3600.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -77,7 +77,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" } diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json index 6cb9e4c68657a96730992cc792db40175e365b88..fafcfff815e7a0f4bf4aebc837c3736145f4165b 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json @@ -7,49 +7,49 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.cap3600.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.pre_framealign.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/summary.jsonl" } ], @@ -82,7 +82,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" } diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json index dac460ed04191b1cbd556f4f34288ef5335993c0..b22b174563e27f7f7cdb0697dd37133f20af601f 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/checkpoints/steps_7000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "demon_attack_fix_latency_8_200ep", diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json index b92aa25c51f18fc3ea3df380eb3377b9fa57eeeb..e247a45ec933341b394d761be6a0b2df5b536ec8 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json @@ -7,44 +7,44 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/post_train/step_5000.json" } ], @@ -65,7 +65,7 @@ "training_data_mix": "demon_attack_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -111,7 +111,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null } @@ -120,7 +120,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json index da4490b149710240d900890004bf8d313790ab9f..1fe5e15832f1b93972e4c268029258603aa6aa9b 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json @@ -7,7 +7,7 @@ "checkpoint": [], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2" }, "evaluation": [], @@ -76,7 +76,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2" } ] diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json index 78deb083c9c42c91ba124d59d2d175134d53c979..6f2691c9c51019c420b999b6eadf76df347b6ca2 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json @@ -7,54 +7,54 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.max3600.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part0of2.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part1of2.json" } ], @@ -130,7 +130,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2" } ] diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json index 2958706551586e40a3f8440102199a46c1e01bf6..465b51df4ab365ddadcf249ffb811d4992b3052b 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json @@ -7,44 +7,44 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json" } ], @@ -65,7 +65,7 @@ "training_data_mix": "demon_attack_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -111,7 +111,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null } @@ -120,7 +120,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json index 3b5f7474324239e1c67dc443cc60a45fc44bfe4b..8411ea713e30f05fe50d4c5261f7fed8aae02b93 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json @@ -7,44 +7,44 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json" } ], @@ -65,7 +65,7 @@ "training_data_mix": "demon_attack_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -111,7 +111,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15", "source_subdir": null } @@ -120,7 +120,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json index 5e3eca87ecb0e2c0275da7e5712833babb8fd9f6..2c779240c97e0dae29a210fd213c18ca8c8ab84a 100644 --- a/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_2000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small" }, "evaluation": [], diff --git a/latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json b/latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json index cc8304a38d6a05ddf2d7aa6263c206dce7cea38b..32eaedbddb776996e61de43985b57041e63366e1 100644 --- a/latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json +++ b/latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json @@ -7,49 +7,49 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/eval_latency_6/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/queue_eval_latency_6.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/eval/fixed_6/queue_eval_results.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/summary.jsonl" } ], @@ -82,7 +82,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" } diff --git a/latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json b/latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json index 6f1f43602b9c1f3172542c6e8f000bab4a9e279d..3d94dd2edcc5a1070fb4df3b37fad2298c2813d4 100644 --- a/latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json +++ b/latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json @@ -7,29 +7,29 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/checkpoint_p0/checkpoint_000048840_25034752.pth" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/git.diff" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/sf_log.txt" } ], diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json index e52777284d222a3ce4dfc382d94e75fa9daf0868..affef91b9e964b4cc53f354070f66397f46eae20 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_curriculum_cumulative_40ep_per_latency", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json index 509594057c3dcf1c5fa51b7db16cb5de79cc3dcb..3686faf93838058977f2c0391e405e24c9b8ed43 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_curriculum_exclusive_40ep_per_latency", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json index 425c87fd13aa0b0da41a067b1de8420e3f393257..3f59220f7657940f7917bca799e967f7442159dc 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep" }, "evaluation": [], @@ -75,12 +75,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_fix_latency_1_200ep", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json index 00f7833b9d0f634f39cc68168ff9841beb655749..e2229f0f0da04c2426d05a68e309bf1e77732029 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep" }, "evaluation": [], @@ -76,12 +76,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_fix_latency_2_200ep", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json index 2815eca9bc6de0440e0d5f86e5f99c5bdc819e28..2f35cf13461a444e1beb00ce5636b777159bc3de 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json @@ -7,204 +7,204 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/checkpoints/steps_4000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/summary.jsonl" } ], @@ -237,7 +237,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps", "config_name": "flappy_fix_latency_2_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json index fe6837bc4addef70048fd5799ecf0f42bd55e284..84f78347b19a0de38a245880b59d1cae868ef853 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps", "config_name": "flappy_fix_latency_2_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json index faa48f473a156a4d6a1a72e9f2e38221b46cf1ec..d11a77cef3897980d01f0ac14175776f39e3dfeb 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps", "config_name": "flappy_fix_latency_2_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json index afdc6b4780c80ab15358bcfc1c93f23a66db417f..bbbd19b9b99d6e93321b80b5b17b3b35b974ae62 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps", "config_name": "flappy_fix_latency_2_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json index 9353d5f4b3f855af158ac19d08bd7cc7f534bd47..c0fa0a07bbe3e909e043b76e03691eff09f2d3c1 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps", "config_name": "flappy_fix_latency_2_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json index 973dbc6cdc0c2a7174197ca44d06b6e7b7788eb1..67993a7b26a226fd29fdf38e50c6c3f5230a8e29 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep" }, "evaluation": [], @@ -76,12 +76,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_fix_latency_3_200ep", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json index b0a33c80a2e7d215b999081a467c141644e368c6..22217a12374cd0eccb4321172aeb04077fde9ea7 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json index 61b8b1a21f249d20797f736a5bdabf28d4ab0682..01003fb9156f54cbc5116fa82559a23c5e1c3338 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json index 74d172cd2780e078d94a3ba6e8fc1c19a3f1989f..3260d127b0d3407a29e9032b8a9524a78fa2e55f 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json index 2324d74a5fdee9790d245551cf36e6a03d51b936..7de4185a8a34a312a098c7c8071575f3e553e3cb 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json @@ -7,384 +7,384 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval_inputs/flappy_fixed_latency_3.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/summary.jsonl" } ], @@ -417,7 +417,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json index c18fc33b3138dc55de91d21679faec67a11310a4..125195c6aa3ea9aeba5d57c16cc0964401447dc2 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep" }, "evaluation": [], @@ -76,12 +76,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_fix_latency_4_200ep", diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json index 7f447b852ed9062d1a17939cdfbc07683cb6f70f..093cf8e07bffad2ea1f812f4205ac857cf52e6b1 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/eval/post_train/step_5000.json" } ], @@ -40,7 +40,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -86,7 +86,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null } @@ -95,7 +95,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json index cd46e0e638ce1aa9195c229ed8015b57212dd4ce..19605e178ef425508ee7bdfcce71e389ae282a87 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json" } ], @@ -40,7 +40,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -86,7 +86,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null } @@ -95,7 +95,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json index 07879af559ed70d24d7a40e205ea5e66030ee7ce..e7aca856d0f9de1e65751ed56668299f575fec89 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json" } ], @@ -40,7 +40,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null }, @@ -86,7 +86,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15", "source_subdir": null } @@ -95,7 +95,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15" } ] diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json index 86fae2f45b1bc64291c200b8c6750da74ffe516e..be4d85359feca0a93f325d0cbbe9e5e81569bc41 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/eval/post_train/step_4000.json" } ], @@ -95,7 +95,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost7" } ] diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json index da354c33b9755c42766c6b4524d02fa29608983a..91906beea5f7c9cde20777d3fa2e9ef050dbf211 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json @@ -2,12 +2,12 @@ "condition": "with_latency_prompt", "checkpoint": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/checkpoints/model.safetensors" }, "evaluation": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2/eval/post_train/step_5000.json" }, "source": { diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json index 0a738cd8ba5d7bf7b2349aa086e1dad5e7b28b1a..47fc858f08b57f34efd160887e720aec9e3977ac 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json @@ -2,12 +2,12 @@ "condition": "no_latency_information", "checkpoint": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/checkpoints/model.safetensors" }, "evaluation": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/eval/post_train/step_5000.json" }, "source": { diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json index 90bb0bba71df900712ba610518768b027d5ab1e6..23f4a2e0d9ae0ccd2eb354d5fc5541011afe2f65 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_1_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_1_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json index 1119b5ade22569105e968bbba9875348896f9840..0a917220655f2013a23173435758aefc9917e7d0 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp" }, "evaluation": [], diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json index 7c95caa12241e2cca7b60dea9ea44d7058f8b021..0ed3c39ff0b0c8ba44014e9892e237d5cee21c4d 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2" }, "evaluation": [], diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json index add775d2f0480c1dccd662d40edc7bae98d786ae..8ef7621b66634067113ca72758db84e3debad58f 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json index 36a104f5c23a84c6075fc53035249f9b5bfb5ef7..7153e2d7f0b6693b70ad3563027e91013c483359 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_3_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_3_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/artifact.json index b73fd808154d0f90c4d630621d604501d9b2a36e..a8f2076d73eec1a23b0c4284759337e10b6ab546 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/checkpoints/steps_2500_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp/checkpoints/steps_2500_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_exp" }, "evaluation": [], diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/artifact.json index 1f0ecdd119723dd0b81afb49d4b377ba2984d55d..ad1d31a36084dbabaa65ea37e22fcb659b57a2ce 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_4_small" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_4_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_4_small", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/artifact.json b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/artifact.json index fe12c83502315d668d99d1b323736169abaf6c96..8dd461fb66e88e8f28007621fe475452c5c08033 100644 --- a/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/artifact.json +++ b/latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/checkpoints/steps_5000_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_6" }, "evaluation": [], diff --git a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/artifact.json b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/artifact.json index c1135e8d097e63250ace0ac70ec9bcbac4ea652e..814b124e516011c04c7f7674787677b6d38c5afc 100644 --- a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/artifact.json +++ b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_1_context5_2000_effbs128" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_1_context5_2000_effbs128/eval/post_train/step_2000.json" } ], @@ -60,7 +60,7 @@ "source_model_revision": "1ba5d2f489f03f9cb07a637ca680fa4f2f8e51a4", "dataset": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_1_200ep_context5", "source_repo": "latency-sensitive-bench/flappy_200ep_context5", "source_revision": "19d800e5f7ec9d5311f508dbfe4a50463261616c", @@ -78,7 +78,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_1_200ep_context5" } ] diff --git a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/artifact.json b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/artifact.json index 7eee79c3810c6ff54c19bbcaa2d9a4f9c7136947..aefd155faae9ef596872f92e9482cde0e61ad49a 100644 --- a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/artifact.json +++ b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_2_context5_2000_effbs128" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_2_context5_2000_effbs128/eval/post_train/step_2000.json" } ], @@ -40,7 +40,7 @@ "source_model_revision": "1ba5d2f489f03f9cb07a637ca680fa4f2f8e51a4", "dataset": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_context5", "source_repo": "latency-sensitive-bench/flappy_200ep_context5", "source_revision": "19d800e5f7ec9d5311f508dbfe4a50463261616c", @@ -60,7 +60,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_2_200ep_context5" } ] diff --git a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/artifact.json b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/artifact.json index 590040f632d609ff8a53ebbbd4a9a7559d91920d..87f1753913cf81410eec5cca7dbe434cde10ef90 100644 --- a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/artifact.json +++ b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_2000_effbs128" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_2000_effbs128/eval/post_train/step_2000.json" } ], @@ -60,7 +60,7 @@ "source_model_revision": "1ba5d2f489f03f9cb07a637ca680fa4f2f8e51a4", "dataset": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_3_200ep_context5", "source_repo": "latency-sensitive-bench/flappy_200ep_context5", "source_revision": "19d800e5f7ec9d5311f508dbfe4a50463261616c", @@ -78,7 +78,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_3_200ep_context5" } ] diff --git a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/artifact.json b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/artifact.json index c14aae0e728e0a81383afa16174015233a2c2aa2..7b743e622433c1fd1a8e388d4e717c8e06f18f1b 100644 --- a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/artifact.json +++ b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/artifact.json @@ -7,29 +7,29 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_latency_3.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_results.jsonl" } ], @@ -62,7 +62,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" } diff --git a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/artifact.json b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/artifact.json index d6330e36836799409c038596ad94068050027008..4824dfeca74e2277f54a6ee6730b60f2fd471720 100644 --- a/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/artifact.json +++ b/latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/artifact.json @@ -7,39 +7,39 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_4_context5_2000_effbs128" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_4_context5_2000_effbs128/eval/post_train/step_2000.json" } ], @@ -60,7 +60,7 @@ "source_model_revision": "1ba5d2f489f03f9cb07a637ca680fa4f2f8e51a4", "dataset": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_4_200ep_context5", "source_repo": "latency-sensitive-bench/flappy_200ep_context5", "source_revision": "19d800e5f7ec9d5311f508dbfe4a50463261616c", @@ -78,7 +78,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "latency-aware/flappy/flappy_fix_latency_4_200ep_context5" } ] diff --git a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/artifact.json b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/artifact.json index c4bd8af96d9e92f307a6ae5954b34b59cef7036e..4b5bb6cd8188ac1663d9eed1fb6f9a81c0028ef7 100644 --- a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/artifact.json +++ b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/artifact.json @@ -7,64 +7,64 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/eval/post_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/resolved_config.yaml" } ], @@ -85,7 +85,7 @@ "training_data_mix": "openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp2", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" }, @@ -135,7 +135,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" } diff --git a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/artifact.json b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/artifact.json index 50223f24723222920dd711a10706bb5dd0e4e2c8..2ed02f7ae51a5520c93f86540e6f04a37cd2b638 100644 --- a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/artifact.json +++ b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/artifact.json @@ -7,64 +7,64 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/eval/post_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/resolved_config.yaml" } ], @@ -85,7 +85,7 @@ "training_data_mix": "openvla_inverted_pendulum_mixed_latency_024_og_l0_native_exp3", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" }, @@ -135,7 +135,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" } diff --git a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/artifact.json b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/artifact.json index 3364c66119b061e5dca822fb758b629e39571968..029f0d0465d43958bbee2b0267677bcd0247ab79 100644 --- a/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/artifact.json +++ b/latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/artifact.json @@ -7,64 +7,64 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "latency-aware/inverted-pendulum/vla/starvla-qwenoft-h1/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/eval/post_train/step_5000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/resolved_config.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/observation_attempts.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/inverted-pendulum/openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2/latency_bench_eval/post_train/step_5000/gymnasium/latency_4/resolved_config.yaml" } ], @@ -85,7 +85,7 @@ "training_data_mix": "openvla_inverted_pendulum_mixed_latency_024_og_l0_native_no_latency_prompt_exp2", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" }, @@ -135,7 +135,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "source_subdir": "inverted_pendulum_fixed_latency_0_100ep_1ksteps_og,inverted_pendulum_fixed_latency_2_1000ep_7k2steps,inverted_pendulum_fixed_latency_4_1000ep_7k2steps" } diff --git a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/artifact.json b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/artifact.json index c7f33c06d31c7fd86ffa7ad69f9b58a53f57cd03..f97c299fe95bf6449ed5f804723e9d51449950a0 100644 --- a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/artifact.json +++ b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/artifact.json @@ -7,34 +7,34 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -67,7 +67,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "zero-latency/deadly-corridor/deadly_corridor_fixed_latency_0_1000ep_7k2steps", "config_name": "deadly_corridor_fixed_latency_0_1000ep_7k2steps" } diff --git a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/artifact.json b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/artifact.json index 00076fd66418a348bbfa4fe3b0d37e7895ce7228..8b9a1055e517cb62ef3cbea1d3a83e5af4951e81 100644 --- a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/artifact.json +++ b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0/checkpoints/steps_500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "deadly_corridor_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/artifact.json b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/artifact.json index 4338fcac092b32986ade311f338880efb7e719b7..abda22221ce22c1f0c26a0767fec72310fadbbdb 100644 --- a/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/artifact.json +++ b/zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full/checkpoints/steps_2500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_0_full" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "deadly_corridor_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -82,7 +82,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep/artifact.json b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep/artifact.json index b1d5ee307d3896df1cdb7967c5aadcf1544e0772..ee01dbc8c82f4ff11533eecefecb6a5e3071d973 100644 --- a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep/artifact.json +++ b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep/checkpoints/steps_7000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep" }, "evaluation": [], @@ -77,12 +77,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "demon_attack_fix_latency_0_200ep", diff --git a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json index b7a6af675ae354b585050420c42a90237148028f..ce6bf963493dd8c45798dc1e30910cec1a129a78 100644 --- a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json +++ b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json @@ -7,34 +7,34 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/demon-attack/demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -67,7 +67,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "zero-latency/demon-attack/demon_attack_fixed_latency_0_200ep_7k2steps", "config_name": "demon_attack_fixed_latency_0_200ep_7k2steps" } diff --git a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small/artifact.json b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small/artifact.json index dadfbf93b30da75149c51a96fc75c9ae757f695b..c3d1196dc8e4b020639e604a3ed3c5e75ecc8d62 100644 --- a/zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small/artifact.json +++ b/zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small/checkpoints/best_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_0_small" }, "evaluation": [], @@ -34,7 +34,7 @@ "training_data_mix": "demon_attack_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -71,7 +71,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep/artifact.json b/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep/artifact.json index c65d17cb7b4f93a7ab5a9efba975b561a35e1728..0196d03fa6640e94f01e2f6cb7efb22372d23df9 100644 --- a/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep/artifact.json +++ b/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep/artifact.json @@ -7,13 +7,13 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep/checkpoints/steps_5000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep" }, "evaluation": [], @@ -75,12 +75,12 @@ "evaluation_evidence": { "episodes": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv" }, "source_manifest": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv" }, "checkpoint_run": "flappy_fix_latency_0_200ep", diff --git a/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json b/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json index c0729a7ab43d0d90b681f9dca1f295cc806c3b48..71d4871e7964cf8daf85c0d6ee21d010bdad6177 100644 --- a/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json +++ b/zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline/artifact.json @@ -7,374 +7,374 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_0_200ep_7k2steps_single_baseline" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_0/episode_metrics.jsonl" }, { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/flappy_fix_latency_0_200ep_7k2steps_single_baseline/summary.jsonl" } ], @@ -407,7 +407,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "zero-latency/flappy/flappy_fixed_latency_0_200ep_7k2steps", "config_name": "flappy_fixed_latency_0_200ep_7k2steps" } diff --git a/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/artifact.json b/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/artifact.json index 6aecb1c690db9613cf679d5622e88ad44615e8de..c6bfbff3bb008f4908bb745400b04c8a8df2b601 100644 --- a/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/artifact.json +++ b/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/checkpoints/steps_5000_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0/checkpoints/steps_5000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0" }, "evaluation": [], diff --git a/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/artifact.json b/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/artifact.json index cbd5b345240c3e644772f6d188595dc5f0189a46..d1249be3af6915743bfffa51d50a17a5b410e02a 100644 --- a/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/artifact.json +++ b/zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/artifact.json @@ -7,18 +7,18 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/checkpoints/best_model.safetensors" }, { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small/checkpoints/steps_1500_model.safetensors" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_0_small" }, "evaluation": [], @@ -39,7 +39,7 @@ "training_data_mix": "flappy_train__bridge", "dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." }, @@ -83,7 +83,7 @@ "interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.", "training_recipe_dataset_reference": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "", "evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately." } diff --git a/zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/artifact.json b/zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/artifact.json index e82b39ce7a2ca42202791d1035bd41812d8d5777..5fa7e545abdf04b93f317e800c56448c3a08a891 100644 --- a/zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/artifact.json +++ b/zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/artifact.json @@ -7,19 +7,19 @@ "checkpoint": [ { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/checkpoints/steps_2000_pytorch_model.pt" } ], "metadata": { "repo": "latency-sensitive-bench/benchmark-models", - "revision": "403deb18146f6780ac28536f78167b21aa18898b", + "revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c", "path": "zero-latency/flappy/vla/starvla-wanoft-h8/wan_oft_flappy_fix_latency_0_context5_2000_effbs128" }, "evaluation": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "evaluations/flappy/wan_oft_flappy_fix_latency_0_context5_2000_effbs128/eval/post_train/step_2000.json" } ], @@ -40,7 +40,7 @@ "source_model_revision": "1ba5d2f489f03f9cb07a637ca680fa4f2f8e51a4", "dataset": { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "zero-latency/flappy/flappy_fix_latency_0_200ep_context5", "source_repo": "latency-sensitive-bench/flappy_200ep_context5", "source_revision": "19d800e5f7ec9d5311f508dbfe4a50463261616c", @@ -58,7 +58,7 @@ "training_data": [ { "repo": "latency-sensitive-bench/benchmark-datasets", - "revision": "27d2fe8badf83302809f5570e3118bd987f15600", + "revision": "a342eb8d051c0982b6af4f80407faad1a2a501db", "path": "zero-latency/flappy/flappy_fix_latency_0_200ep_context5" } ]