Add concise experiment navigation and task-level reproducibility manifests
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json +5 -5
- latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json +8 -8
- latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json +8 -8
- latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json +10 -10
- latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json +8 -8
- latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json +5 -5
- latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json +6 -6
- latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json +5 -5
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json +4 -4
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json +4 -4
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json +4 -4
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json +9 -9
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json +10 -10
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json +9 -9
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json +10 -10
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json +4 -4
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json +11 -11
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json +2 -2
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json +11 -11
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json +11 -11
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json +11 -11
- latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json +2 -2
- latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json +10 -10
- latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json +5 -5
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json +41 -41
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json +75 -75
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json +77 -77
- latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json +6 -6
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json +6 -6
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json +6 -6
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json +4 -4
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json +2 -2
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json +2 -2
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json +5 -5
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json +3 -3
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json +3 -3
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json +5 -5
- latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json +5 -5
latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/artifact.json
CHANGED
|
@@ -7,29 +7,29 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/small-policy/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/artifact.json
CHANGED
|
@@ -7,39 +7,39 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
@@ -72,7 +72,7 @@
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
-
"revision": "
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/artifact.json
CHANGED
|
@@ -7,39 +7,39 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
@@ -72,7 +72,7 @@
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
-
"revision": "
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/artifact.json
CHANGED
|
@@ -7,49 +7,49 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_2000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval_inputs/deadly_corridor_fixed_latency_6.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
@@ -82,7 +82,7 @@
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 87 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 88 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_2000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/eval_inputs/deadly_corridor_fixed_latency_6.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 87 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 88 |
}
|
latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/artifact.json
CHANGED
|
@@ -7,39 +7,39 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
@@ -72,7 +72,7 @@
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
-
"revision": "
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/deadly-corridor/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/summary.jsonl"
|
| 44 |
}
|
| 45 |
],
|
|
|
|
| 72 |
"training_data": [
|
| 73 |
{
|
| 74 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 75 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 76 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 77 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 78 |
}
|
latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/steps_500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
@@ -39,7 +39,7 @@
|
|
| 39 |
"training_data_mix": "deadly_corridor_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
@@ -83,7 +83,7 @@
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
-
"revision": "
|
| 87 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2/checkpoints/steps_500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/deadly-corridor/vla/starvla-qwenoft-h1/openvla_deadly_corridor_fix_latency_2"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 39 |
"training_data_mix": "deadly_corridor_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 87 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fix_latency_2",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/artifact.json
CHANGED
|
@@ -7,29 +7,29 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
|
@@ -62,7 +62,7 @@
|
|
| 62 |
"training_data": [
|
| 63 |
{
|
| 64 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 65 |
-
"revision": "
|
| 66 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 67 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 68 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/deadly-corridor/vla/starvla-wanoft-h8/wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
|
|
|
| 62 |
"training_data": [
|
| 63 |
{
|
| 64 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 65 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 66 |
"path": "latency-aware/deadly-corridor/deadly_corridor_fixed_latency_6_1000ep_7k2steps",
|
| 67 |
"config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps"
|
| 68 |
}
|
latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/artifact.json
CHANGED
|
@@ -7,29 +7,29 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/small-policy/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_2_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_2_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_2_200ep",
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_4_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_4_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_4_200ep",
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_6_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_6_200ep",
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/artifact.json
CHANGED
|
@@ -7,44 +7,44 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 49 |
}
|
| 50 |
],
|
|
@@ -77,7 +77,7 @@
|
|
| 77 |
"training_data": [
|
| 78 |
{
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 82 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 83 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 49 |
}
|
| 50 |
],
|
|
|
|
| 77 |
"training_data": [
|
| 78 |
{
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 82 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 83 |
}
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/artifact.json
CHANGED
|
@@ -7,49 +7,49 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.pre_framealign.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
@@ -82,7 +82,7 @@
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/eval/post_train/step_4000.pre_framealign.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/artifact.json
CHANGED
|
@@ -7,44 +7,44 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 49 |
}
|
| 50 |
],
|
|
@@ -77,7 +77,7 @@
|
|
| 77 |
"training_data": [
|
| 78 |
{
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 82 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 83 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 49 |
}
|
| 50 |
],
|
|
|
|
| 77 |
"training_data": [
|
| 78 |
{
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 82 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 83 |
}
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/artifact.json
CHANGED
|
@@ -7,49 +7,49 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.pre_framealign.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
@@ -82,7 +82,7 @@
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_6_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.cap3600.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/eval/post_train/step_4000.pre_framealign.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/demon-attack/demon_attack_fix_latency_6_200ep_7k2steps_stitch/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_8_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep/checkpoints/steps_7000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/demon_attack_fix_latency_8_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/demon-attack/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "demon_attack_fix_latency_8_200ep",
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/artifact.json
CHANGED
|
@@ -7,44 +7,44 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
@@ -65,7 +65,7 @@
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
-
"revision": "
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
@@ -111,7 +111,7 @@
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
-
"revision": "
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
@@ -120,7 +120,7 @@
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
-
"revision": "
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2/artifact.json
CHANGED
|
@@ -7,7 +7,7 @@
|
|
| 7 |
"checkpoint": [],
|
| 8 |
"metadata": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 12 |
},
|
| 13 |
"evaluation": [],
|
|
@@ -76,7 +76,7 @@
|
|
| 76 |
"training_data": [
|
| 77 |
{
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
-
"revision": "
|
| 80 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 81 |
}
|
| 82 |
]
|
|
|
|
| 7 |
"checkpoint": [],
|
| 8 |
"metadata": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 12 |
},
|
| 13 |
"evaluation": [],
|
|
|
|
| 76 |
"training_data": [
|
| 77 |
{
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 80 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 81 |
}
|
| 82 |
]
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/artifact.json
CHANGED
|
@@ -7,54 +7,54 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.max3600.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part0of2.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part1of2.json"
|
| 59 |
}
|
| 60 |
],
|
|
@@ -130,7 +130,7 @@
|
|
| 130 |
"training_data": [
|
| 131 |
{
|
| 132 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 133 |
-
"revision": "
|
| 134 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 135 |
}
|
| 136 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.max3600.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part0of2.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2_exp3/eval/post_train/step_4000.part1of2.json"
|
| 59 |
}
|
| 60 |
],
|
|
|
|
| 130 |
"training_data": [
|
| 131 |
{
|
| 132 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 133 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 134 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_2"
|
| 135 |
}
|
| 136 |
]
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/artifact.json
CHANGED
|
@@ -7,44 +7,44 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
@@ -65,7 +65,7 @@
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
-
"revision": "
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
@@ -111,7 +111,7 @@
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
-
"revision": "
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
@@ -120,7 +120,7 @@
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
-
"revision": "
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/artifact.json
CHANGED
|
@@ -7,44 +7,44 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
@@ -65,7 +65,7 @@
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
-
"revision": "
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
@@ -111,7 +111,7 @@
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
-
"revision": "
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
@@ -120,7 +120,7 @@
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
-
"revision": "
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_2000.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_3000.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_4000.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/mid_train/step_5000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/openvla_bridge_demon_attack_fixed_latency_6_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json"
|
| 49 |
}
|
| 50 |
],
|
|
|
|
| 65 |
"training_data_mix": "demon_attack_train__bridge",
|
| 66 |
"dataset_reference": {
|
| 67 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 68 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 69 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 70 |
"source_subdir": null
|
| 71 |
},
|
|
|
|
| 111 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 112 |
"training_recipe_dataset_reference": {
|
| 113 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 114 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 115 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15",
|
| 116 |
"source_subdir": null
|
| 117 |
}
|
|
|
|
| 120 |
"training_data": [
|
| 121 |
{
|
| 122 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 123 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 124 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps_ghost15"
|
| 125 |
}
|
| 126 |
]
|
latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_2000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small/checkpoints/steps_2000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-qwenoft-h1/openvla_demon_attack_fix_latency_2_small"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/artifact.json
CHANGED
|
@@ -7,49 +7,49 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"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"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
@@ -82,7 +82,7 @@
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/demon-attack/vla/starvla-wanoft-h8/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"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"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/hydra.yaml"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/.hydra/overrides.yaml"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/hydra/train_starvla_hydra.log"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/demon-attack/wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/summary.jsonl"
|
| 54 |
}
|
| 55 |
],
|
|
|
|
| 82 |
"training_data": [
|
| 83 |
{
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "latency-aware/demon-attack/demon_attack_fixed_latency_6_200ep_7k2steps",
|
| 87 |
"config_name": "demon_attack_fixed_latency_6_200ep_7k2steps"
|
| 88 |
}
|
latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/artifact.json
CHANGED
|
@@ -7,29 +7,29 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"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"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/small-policy/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"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"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"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"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"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"
|
| 34 |
}
|
| 35 |
],
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "flappy_curriculum_cumulative_40ep_per_latency",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_cumulative_40ep_per_latency"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "flappy_curriculum_cumulative_40ep_per_latency",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -77,12 +77,12 @@
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
-
"revision": "
|
| 81 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
-
"revision": "
|
| 86 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "flappy_curriculum_exclusive_40ep_per_latency",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_curriculum_exclusive_40ep_per_latency"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 77 |
"evaluation_evidence": {
|
| 78 |
"episodes": {
|
| 79 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 80 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 81 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 82 |
},
|
| 83 |
"source_manifest": {
|
| 84 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 85 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 86 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 87 |
},
|
| 88 |
"checkpoint_run": "flappy_curriculum_exclusive_40ep_per_latency",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -75,12 +75,12 @@
|
|
| 75 |
"evaluation_evidence": {
|
| 76 |
"episodes": {
|
| 77 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 78 |
-
"revision": "
|
| 79 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 80 |
},
|
| 81 |
"source_manifest": {
|
| 82 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 83 |
-
"revision": "
|
| 84 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 85 |
},
|
| 86 |
"checkpoint_run": "flappy_fix_latency_1_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_1_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 75 |
"evaluation_evidence": {
|
| 76 |
"episodes": {
|
| 77 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 78 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 79 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 80 |
},
|
| 81 |
"source_manifest": {
|
| 82 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 83 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 84 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 85 |
},
|
| 86 |
"checkpoint_run": "flappy_fix_latency_1_200ep",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -76,12 +76,12 @@
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
-
"revision": "
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
-
"revision": "
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_2_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_2_200ep",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/artifact.json
CHANGED
|
@@ -7,204 +7,204 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/checkpoints/steps_4000_pytorch_model.pt"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/summary.jsonl"
|
| 209 |
}
|
| 210 |
],
|
|
@@ -237,7 +237,7 @@
|
|
| 237 |
"training_data": [
|
| 238 |
{
|
| 239 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 240 |
-
"revision": "
|
| 241 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 242 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 243 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory/checkpoints/steps_4000_pytorch_model.pt"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory/summary.jsonl"
|
| 209 |
}
|
| 210 |
],
|
|
|
|
| 237 |
"training_data": [
|
| 238 |
{
|
| 239 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 240 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 241 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 242 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 243 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"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"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"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"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"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"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"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"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"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"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"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"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"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"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"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"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"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"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"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"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"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"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"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"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"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"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"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"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"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"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"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"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"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"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"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"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"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"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"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"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"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"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"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"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"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"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"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"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"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"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"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"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"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"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"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"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"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"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"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"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"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"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"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"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"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"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"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"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_2_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_2/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_2/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_2/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_2/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_2/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_2/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_2/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_2/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_2/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_2/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_2/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_2/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_2/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_2/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_2/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_2/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_2_200ep_7k2steps_stitch/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fix_latency_2_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fix_latency_2_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -76,12 +76,12 @@
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
-
"revision": "
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
-
"revision": "
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_3_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_3_200ep",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"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"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"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"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"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"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"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"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"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"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"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"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"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"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"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"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"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"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"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"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"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"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"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"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"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"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"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"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"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"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"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"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"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"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"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"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"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"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"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"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"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"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"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"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"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"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"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"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"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"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"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"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"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"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"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"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"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"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"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"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"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"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"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"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"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"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"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"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_plain_multi"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_plain_multi/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/artifact.json
CHANGED
|
@@ -7,374 +7,374 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
@@ -407,7 +407,7 @@
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
-
"revision": "
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_single_baseline"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/eval/post_train/step_4000.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/hydra.yaml"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/.hydra/overrides.yaml"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/hydra/train_starvla_hydra.log"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_single_baseline/summary.jsonl"
|
| 379 |
}
|
| 380 |
],
|
|
|
|
| 407 |
"training_data": [
|
| 408 |
{
|
| 409 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 410 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 411 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 412 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 413 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/artifact.json
CHANGED
|
@@ -7,384 +7,384 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
-
"revision": "
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
-
"revision": "
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
-
"revision": "
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
-
"revision": "
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
-
"revision": "
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
-
"revision": "
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
-
"revision": "
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
-
"revision": "
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
-
"revision": "
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
-
"revision": "
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
-
"revision": "
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
-
"revision": "
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
-
"revision": "
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
-
"revision": "
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
-
"revision": "
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_1500.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
-
"revision": "
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
-
"revision": "
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval_inputs/flappy_fixed_latency_3.json"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
-
"revision": "
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
-
"revision": "
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
-
"revision": "
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
-
"revision": "
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
-
"revision": "
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
-
"revision": "
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
-
"revision": "
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
-
"revision": "
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
-
"revision": "
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
-
"revision": "
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
-
"revision": "
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
-
"revision": "
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
-
"revision": "
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
-
"revision": "
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
-
"revision": "
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
-
"revision": "
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
-
"revision": "
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
-
"revision": "
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
-
"revision": "
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
-
"revision": "
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
-
"revision": "
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
-
"revision": "
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
-
"revision": "
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
-
"revision": "
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
-
"revision": "
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
-
"revision": "
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
-
"revision": "
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
-
"revision": "
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
-
"revision": "
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
-
"revision": "
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
-
"revision": "
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
-
"revision": "
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
-
"revision": "
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
-
"revision": "
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
-
"revision": "
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
-
"revision": "
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
-
"revision": "
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
-
"revision": "
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
-
"revision": "
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
-
"revision": "
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
-
"revision": "
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
-
"revision": "
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
-
"revision": "
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
-
"revision": "
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
-
"revision": "
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
-
"revision": "
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
-
"revision": "
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
-
"revision": "
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
-
"revision": "
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
-
"revision": "
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
-
"revision": "
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
-
"revision": "
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
-
"revision": "
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 379 |
},
|
| 380 |
{
|
| 381 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 382 |
-
"revision": "
|
| 383 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 384 |
},
|
| 385 |
{
|
| 386 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 387 |
-
"revision": "
|
| 388 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/summary.jsonl"
|
| 389 |
}
|
| 390 |
],
|
|
@@ -417,7 +417,7 @@
|
|
| 417 |
"training_data": [
|
| 418 |
{
|
| 419 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 420 |
-
"revision": "
|
| 421 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 422 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 423 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_3_200ep_7k2steps_stitch"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1000.json"
|
| 24 |
},
|
| 25 |
{
|
| 26 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 27 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 28 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1250.json"
|
| 29 |
},
|
| 30 |
{
|
| 31 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 32 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 33 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1500.json"
|
| 34 |
},
|
| 35 |
{
|
| 36 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 37 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 38 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_1750.json"
|
| 39 |
},
|
| 40 |
{
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2000.json"
|
| 44 |
},
|
| 45 |
{
|
| 46 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 47 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 48 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2250.json"
|
| 49 |
},
|
| 50 |
{
|
| 51 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 52 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 53 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_250.json"
|
| 54 |
},
|
| 55 |
{
|
| 56 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 57 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 58 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2500.json"
|
| 59 |
},
|
| 60 |
{
|
| 61 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 62 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 63 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_2750.json"
|
| 64 |
},
|
| 65 |
{
|
| 66 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 67 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 68 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3000.json"
|
| 69 |
},
|
| 70 |
{
|
| 71 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 72 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 73 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3250.json"
|
| 74 |
},
|
| 75 |
{
|
| 76 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 77 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 78 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3500.json"
|
| 79 |
},
|
| 80 |
{
|
| 81 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 82 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 83 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_3750.json"
|
| 84 |
},
|
| 85 |
{
|
| 86 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 87 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 88 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_4000.json"
|
| 89 |
},
|
| 90 |
{
|
| 91 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 92 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 93 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_500.json"
|
| 94 |
},
|
| 95 |
{
|
| 96 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 97 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 98 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/mid_train/step_750.json"
|
| 99 |
},
|
| 100 |
{
|
| 101 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 102 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 103 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_1500.json"
|
| 104 |
},
|
| 105 |
{
|
| 106 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 107 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 108 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval/post_train/step_4000.json"
|
| 109 |
},
|
| 110 |
{
|
| 111 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 112 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 113 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/eval_inputs/flappy_fixed_latency_3.json"
|
| 114 |
},
|
| 115 |
{
|
| 116 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 117 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 118 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/hydra.yaml"
|
| 119 |
},
|
| 120 |
{
|
| 121 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 122 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 123 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/.hydra/overrides.yaml"
|
| 124 |
},
|
| 125 |
{
|
| 126 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 127 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 128 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/hydra/train_starvla_hydra.log"
|
| 129 |
},
|
| 130 |
{
|
| 131 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 132 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 133 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_0.json"
|
| 134 |
},
|
| 135 |
{
|
| 136 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 137 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 138 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/_progress/rank_1.json"
|
| 139 |
},
|
| 140 |
{
|
| 141 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 142 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 143 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1000/flappy/latency_3/episode_metrics.jsonl"
|
| 144 |
},
|
| 145 |
{
|
| 146 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 147 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 148 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_0.json"
|
| 149 |
},
|
| 150 |
{
|
| 151 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 152 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 153 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/_progress/rank_1.json"
|
| 154 |
},
|
| 155 |
{
|
| 156 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 157 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 158 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1250/flappy/latency_3/episode_metrics.jsonl"
|
| 159 |
},
|
| 160 |
{
|
| 161 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 162 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 163 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_0.json"
|
| 164 |
},
|
| 165 |
{
|
| 166 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 167 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 168 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/_progress/rank_1.json"
|
| 169 |
},
|
| 170 |
{
|
| 171 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 172 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 173 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1500/flappy/latency_3/episode_metrics.jsonl"
|
| 174 |
},
|
| 175 |
{
|
| 176 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 177 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 178 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_0.json"
|
| 179 |
},
|
| 180 |
{
|
| 181 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 182 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 183 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/_progress/rank_1.json"
|
| 184 |
},
|
| 185 |
{
|
| 186 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 187 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 188 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_1750/flappy/latency_3/episode_metrics.jsonl"
|
| 189 |
},
|
| 190 |
{
|
| 191 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 192 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 193 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_0.json"
|
| 194 |
},
|
| 195 |
{
|
| 196 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 197 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 198 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/_progress/rank_1.json"
|
| 199 |
},
|
| 200 |
{
|
| 201 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 202 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 203 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2000/flappy/latency_3/episode_metrics.jsonl"
|
| 204 |
},
|
| 205 |
{
|
| 206 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 207 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 208 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_0.json"
|
| 209 |
},
|
| 210 |
{
|
| 211 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 212 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 213 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/_progress/rank_1.json"
|
| 214 |
},
|
| 215 |
{
|
| 216 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 217 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 218 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2250/flappy/latency_3/episode_metrics.jsonl"
|
| 219 |
},
|
| 220 |
{
|
| 221 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 222 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 223 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_0.json"
|
| 224 |
},
|
| 225 |
{
|
| 226 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 227 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 228 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/_progress/rank_1.json"
|
| 229 |
},
|
| 230 |
{
|
| 231 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 232 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 233 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_250/flappy/latency_3/episode_metrics.jsonl"
|
| 234 |
},
|
| 235 |
{
|
| 236 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 237 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 238 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_0.json"
|
| 239 |
},
|
| 240 |
{
|
| 241 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 242 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 243 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/_progress/rank_1.json"
|
| 244 |
},
|
| 245 |
{
|
| 246 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 247 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 248 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2500/flappy/latency_3/episode_metrics.jsonl"
|
| 249 |
},
|
| 250 |
{
|
| 251 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 252 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 253 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_0.json"
|
| 254 |
},
|
| 255 |
{
|
| 256 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 257 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 258 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/_progress/rank_1.json"
|
| 259 |
},
|
| 260 |
{
|
| 261 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 262 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 263 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_2750/flappy/latency_3/episode_metrics.jsonl"
|
| 264 |
},
|
| 265 |
{
|
| 266 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 267 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 268 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_0.json"
|
| 269 |
},
|
| 270 |
{
|
| 271 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 272 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 273 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/_progress/rank_1.json"
|
| 274 |
},
|
| 275 |
{
|
| 276 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 277 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 278 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3000/flappy/latency_3/episode_metrics.jsonl"
|
| 279 |
},
|
| 280 |
{
|
| 281 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 282 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 283 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_0.json"
|
| 284 |
},
|
| 285 |
{
|
| 286 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 287 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 288 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/_progress/rank_1.json"
|
| 289 |
},
|
| 290 |
{
|
| 291 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 292 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 293 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3250/flappy/latency_3/episode_metrics.jsonl"
|
| 294 |
},
|
| 295 |
{
|
| 296 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 297 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 298 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_0.json"
|
| 299 |
},
|
| 300 |
{
|
| 301 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 302 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 303 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/_progress/rank_1.json"
|
| 304 |
},
|
| 305 |
{
|
| 306 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 307 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 308 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3500/flappy/latency_3/episode_metrics.jsonl"
|
| 309 |
},
|
| 310 |
{
|
| 311 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 312 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 313 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_0.json"
|
| 314 |
},
|
| 315 |
{
|
| 316 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 317 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 318 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/_progress/rank_1.json"
|
| 319 |
},
|
| 320 |
{
|
| 321 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 322 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 323 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_3750/flappy/latency_3/episode_metrics.jsonl"
|
| 324 |
},
|
| 325 |
{
|
| 326 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 327 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 328 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_0.json"
|
| 329 |
},
|
| 330 |
{
|
| 331 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 332 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 333 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/_progress/rank_1.json"
|
| 334 |
},
|
| 335 |
{
|
| 336 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 337 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 338 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 339 |
},
|
| 340 |
{
|
| 341 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 342 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 343 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_0.json"
|
| 344 |
},
|
| 345 |
{
|
| 346 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 347 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 348 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/_progress/rank_1.json"
|
| 349 |
},
|
| 350 |
{
|
| 351 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 352 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 353 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_500/flappy/latency_3/episode_metrics.jsonl"
|
| 354 |
},
|
| 355 |
{
|
| 356 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 357 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 358 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_0.json"
|
| 359 |
},
|
| 360 |
{
|
| 361 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 362 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 363 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/_progress/rank_1.json"
|
| 364 |
},
|
| 365 |
{
|
| 366 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 367 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 368 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/mid_train/step_750/flappy/latency_3/episode_metrics.jsonl"
|
| 369 |
},
|
| 370 |
{
|
| 371 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 372 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 373 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_0.json"
|
| 374 |
},
|
| 375 |
{
|
| 376 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 377 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 378 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/_progress/rank_1.json"
|
| 379 |
},
|
| 380 |
{
|
| 381 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 382 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 383 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/latency_bench_eval/post_train/step_4000/flappy/latency_3/episode_metrics.jsonl"
|
| 384 |
},
|
| 385 |
{
|
| 386 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 387 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 388 |
"path": "evaluations/flappy/flappy_fix_latency_3_200ep_7k2steps_stitch/summary.jsonl"
|
| 389 |
}
|
| 390 |
],
|
|
|
|
| 417 |
"training_data": [
|
| 418 |
{
|
| 419 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 420 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 421 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps",
|
| 422 |
"config_name": "flappy_fixed_latency_3_200ep_7k2steps"
|
| 423 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/artifact.json
CHANGED
|
@@ -7,13 +7,13 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
@@ -76,12 +76,12 @@
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
-
"revision": "
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
-
"revision": "
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_4_200ep",
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/flappy_fix_latency_4_200ep"
|
| 18 |
},
|
| 19 |
"evaluation": [],
|
|
|
|
| 76 |
"evaluation_evidence": {
|
| 77 |
"episodes": {
|
| 78 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 79 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 80 |
"path": "evaluations/flappy/latency-transfer-qwenoft/episodes.csv"
|
| 81 |
},
|
| 82 |
"source_manifest": {
|
| 83 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 84 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 85 |
"path": "evaluations/flappy/latency-transfer-qwenoft/source_manifest.csv"
|
| 86 |
},
|
| 87 |
"checkpoint_run": "flappy_fix_latency_4_200ep",
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/artifact.json
CHANGED
|
@@ -7,19 +7,19 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
@@ -40,7 +40,7 @@
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
-
"revision": "
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
@@ -86,7 +86,7 @@
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
-
"revision": "
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
@@ -95,7 +95,7 @@
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
-
"revision": "
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/artifact.json
CHANGED
|
@@ -7,19 +7,19 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
@@ -40,7 +40,7 @@
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
-
"revision": "
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
@@ -86,7 +86,7 @@
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
-
"revision": "
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
@@ -95,7 +95,7 @@
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
-
"revision": "
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_exp2/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/artifact.json
CHANGED
|
@@ -7,19 +7,19 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
@@ -40,7 +40,7 @@
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
-
"revision": "
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
@@ -86,7 +86,7 @@
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
-
"revision": "
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
@@ -95,7 +95,7 @@
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
-
"revision": "
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15/checkpoints/steps_5000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost15_seed1/eval/post_train/step_5000.json"
|
| 24 |
}
|
| 25 |
],
|
|
|
|
| 40 |
"training_data_mix": "flappy_train__bridge",
|
| 41 |
"dataset_reference": {
|
| 42 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 43 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 44 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 45 |
"source_subdir": null
|
| 46 |
},
|
|
|
|
| 86 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 87 |
"training_recipe_dataset_reference": {
|
| 88 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 89 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 90 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15",
|
| 91 |
"source_subdir": null
|
| 92 |
}
|
|
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost15"
|
| 100 |
}
|
| 101 |
]
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/artifact.json
CHANGED
|
@@ -7,19 +7,19 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
-
"revision": "
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
-
"revision": "
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/eval/post_train/step_4000.json"
|
| 24 |
}
|
| 25 |
],
|
|
@@ -95,7 +95,7 @@
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
-
"revision": "
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost7"
|
| 100 |
}
|
| 101 |
]
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/checkpoints/steps_4000_model.safetensors"
|
| 12 |
}
|
| 13 |
],
|
| 14 |
"metadata": {
|
| 15 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 16 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 17 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3"
|
| 18 |
},
|
| 19 |
"evaluation": [
|
| 20 |
{
|
| 21 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 22 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 23 |
"path": "evaluations/flappy/openvla_bridge_flappy_fixed_latency_3_200ep_7k2steps_ghost7_exp3/eval/post_train/step_4000.json"
|
| 24 |
}
|
| 25 |
],
|
|
|
|
| 95 |
"training_data": [
|
| 96 |
{
|
| 97 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 98 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 99 |
"path": "latency-aware/flappy/flappy_fixed_latency_3_200ep_7k2steps_ghost7"
|
| 100 |
}
|
| 101 |
]
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/artifact.json
CHANGED
|
@@ -2,12 +2,12 @@
|
|
| 2 |
"condition": "with_latency_prompt",
|
| 3 |
"checkpoint": {
|
| 4 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 5 |
-
"revision": "
|
| 6 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/checkpoints/model.safetensors"
|
| 7 |
},
|
| 8 |
"evaluation": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2/eval/post_train/step_5000.json"
|
| 12 |
},
|
| 13 |
"source": {
|
|
|
|
| 2 |
"condition": "with_latency_prompt",
|
| 3 |
"checkpoint": {
|
| 4 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 5 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 6 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2/checkpoints/model.safetensors"
|
| 7 |
},
|
| 8 |
"evaluation": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 10 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 11 |
"path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2/eval/post_train/step_5000.json"
|
| 12 |
},
|
| 13 |
"source": {
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/artifact.json
CHANGED
|
@@ -2,12 +2,12 @@
|
|
| 2 |
"condition": "no_latency_information",
|
| 3 |
"checkpoint": {
|
| 4 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 5 |
-
"revision": "
|
| 6 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/checkpoints/model.safetensors"
|
| 7 |
},
|
| 8 |
"evaluation": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/eval/post_train/step_5000.json"
|
| 12 |
},
|
| 13 |
"source": {
|
|
|
|
| 2 |
"condition": "no_latency_information",
|
| 3 |
"checkpoint": {
|
| 4 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 5 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 6 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/checkpoints/model.safetensors"
|
| 7 |
},
|
| 8 |
"evaluation": {
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 10 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 11 |
"path": "evaluations/flappy/openvla_bridge_flappy_latency_mixed_exp2_no_latency_information/eval/post_train/step_5000.json"
|
| 12 |
},
|
| 13 |
"source": {
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
@@ -39,7 +39,7 @@
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_1_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
@@ -83,7 +83,7 @@
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
-
"revision": "
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_1_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_1_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_1_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_1_small_tmp"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_pytorch_model.pt"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2/checkpoints/steps_5000_pytorch_model.pt"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
@@ -39,7 +39,7 @@
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_2_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
@@ -83,7 +83,7 @@
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
-
"revision": "
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_2_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_2_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_2_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_2_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/artifact.json
CHANGED
|
@@ -7,18 +7,18 @@
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
-
"revision": "
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
-
"revision": "
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
-
"revision": "
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
@@ -39,7 +39,7 @@
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
-
"revision": "
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_3_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
@@ -83,7 +83,7 @@
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
-
"revision": "
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_3_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|
|
|
|
| 7 |
"checkpoint": [
|
| 8 |
{
|
| 9 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 10 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 11 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/best_model.safetensors"
|
| 12 |
},
|
| 13 |
{
|
| 14 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 15 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 16 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small/checkpoints/steps_1500_model.safetensors"
|
| 17 |
}
|
| 18 |
],
|
| 19 |
"metadata": {
|
| 20 |
"repo": "latency-sensitive-bench/benchmark-models",
|
| 21 |
+
"revision": "46266ba5a184ac622f7fa3386ac689d8d755d64c",
|
| 22 |
"path": "latency-aware/flappy/vla/starvla-qwenoft-h1/openvla_flappy_fix_latency_3_small"
|
| 23 |
},
|
| 24 |
"evaluation": [],
|
|
|
|
| 39 |
"training_data_mix": "flappy_train__bridge",
|
| 40 |
"dataset_reference": {
|
| 41 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 42 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 43 |
"path": "latency-aware/flappy/flappy_fix_latency_3_small",
|
| 44 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 45 |
},
|
|
|
|
| 83 |
"interpretation": "The paper experiment type is known; an exact plotted checkpoint binding is recorded only when its immutable source manifest establishes it.",
|
| 84 |
"training_recipe_dataset_reference": {
|
| 85 |
"repo": "latency-sensitive-bench/benchmark-datasets",
|
| 86 |
+
"revision": "a342eb8d051c0982b6af4f80407faad1a2a501db",
|
| 87 |
"path": "latency-aware/flappy/flappy_fix_latency_3_small",
|
| 88 |
"evidence": "Training config uses the exact source dataset name as its local data-root directory. Original metadata and statistics are retained separately."
|
| 89 |
}
|