[ { "run": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "0557aba023581368ee883695e9535217afb5c126d083aa825e12c797c7bd3ddf", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_0_1000ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "deadly_corridor_fix_latency_0_1000ep_7k2steps_single_baseline" }, { "run": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex", "framework": "QwenOFT", "weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "config_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/config.full.yaml", "normalization_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "f3b8a14cb9feeecdddba3d73f011c49d93034b4d2f20a73acc59b7ea66ba395d", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 8, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "deadly_corridor_fix_latency_6_1000ep_7k2steps_kv_memory_flex" }, { "run": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi", "framework": "QwenOFT", "weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "config_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/config.full.yaml", "normalization_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "d7b311bfda72620b79f10575c5acc9b52d1c722e3ee42b20c2878603797e0530", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "deadly_corridor_fix_latency_6_1000ep_7k2steps_plain_multi" }, { "run": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "1dfd3f99e14743813ff4a86047bc66f0a4e81a6e8a4c5970c617d1813721e121", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "deadly_corridor_fix_latency_6_1000ep_7k2steps_single_baseline" }, { "run": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch", "framework": "QwenOFT", "weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "config_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/config.full.yaml", "normalization_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "6ae6507a0c59ba77f90eb1e823c920092f3ad51ad8bb4da7689ccb48b303bb90", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" }, "image_mode": "stitch", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch" }, { "run": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "4f7856234a1dca6d6c678270ec3958d2c603cee4f58f5f2fd9b138ba074bebcc", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_0_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "demon_attack_fix_latency_0_200ep_7k2steps_single_baseline" }, { "run": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex", "framework": "QwenOFT", "weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "config_path": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/config.full.yaml", "normalization_path": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "eef33f41a8d202aced4f896622ec2b872f253e3b453d2fe2497c66cfb05c2cc4", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 8, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "demon_attack_fix_latency_6_200ep_7k2steps_kv_memory_flex" }, { "run": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi", "framework": "QwenOFT", "weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "config_path": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/config.full.yaml", "normalization_path": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "08ba825529c75663280016705b5583bcd4837fd749a7c9ee4b4eb1f75e266eca", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "demon_attack_fix_latency_6_200ep_7k2steps_plain_multi" }, { "run": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "d5319537cca52cfb0e05705561bb0d883fd3fc6e19314fb386ae3b82659a910c", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "demon_attack_fix_latency_6_200ep_7k2steps_single_baseline" }, { "run": "demon_attack_fix_latency_6_200ep_7k2steps_stitch", "framework": "QwenOFT", "weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "config_path": "demon_attack_fix_latency_6_200ep_7k2steps_stitch/config.full.yaml", "normalization_path": "demon_attack_fix_latency_6_200ep_7k2steps_stitch/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "demon_attack_fix_latency_6_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "5bb27757308c5a506424a7c353c5611c3fd4f6cb319a8f9df0c82f696c6c9db9", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" }, "image_mode": "stitch", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "demon_attack_fix_latency_6_200ep_7k2steps_stitch" }, { "run": "flappy_fix_latency_0_200ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_0_200ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "flappy_fix_latency_0_200ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_0_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "e6df7c64a760d025417d84c2bbc1dfe6e2567790a4c18bacd9e634f81c676eeb", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_0_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_0_200ep_7k2steps_single_baseline" }, { "run": "flappy_fix_latency_2_200ep_7k2steps_kv_memory", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory/checkpoints/steps_4000_pytorch_model.pt", "config_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory/config.full.yaml", "normalization_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory/checkpoints/steps_4000_state/pytorch_model/mp_rank_00_model_states.pt", "source_weight_sha256": "a9b3853e369b9453e194e05846a8be58e62292ce743a75ce7304d0e9068031d2", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fix_latency_2_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 8, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_2_200ep_7k2steps_kv_memory" }, { "run": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/config.full.yaml", "normalization_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "2f1e4d04270984d96f78074912c082fa8275d3680797a7a2eff02666dcfc1f25", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fix_latency_2_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 8, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_2_200ep_7k2steps_kv_memory_flex" }, { "run": "flappy_fix_latency_2_200ep_7k2steps_plain_multi", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_2_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_2_200ep_7k2steps_plain_multi/config.full.yaml", "normalization_path": "flappy_fix_latency_2_200ep_7k2steps_plain_multi/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_2_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "639393e062693dbb9e2bd801367353cd11256a4e8a7419e02186d360b5b249b2", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fix_latency_2_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_2_200ep_7k2steps_plain_multi" }, { "run": "flappy_fix_latency_2_200ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_2_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_2_200ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "flappy_fix_latency_2_200ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_2_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "a92627efd00d16c3a3a392b46cc366ec06956b0e651f81951b1be5435cc75f37", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fix_latency_2_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_2_200ep_7k2steps_single_baseline" }, { "run": "flappy_fix_latency_2_200ep_7k2steps_stitch", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_2_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_2_200ep_7k2steps_stitch/config.full.yaml", "normalization_path": "flappy_fix_latency_2_200ep_7k2steps_stitch/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_2_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "d29f05e6565500aecc785f04b687b107f703021b59947f2d261d167fb932c3f5", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fix_latency_2_200ep_7k2steps" }, "image_mode": "stitch", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_2_200ep_7k2steps_stitch" }, { "run": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/config.full.yaml", "normalization_path": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "4a08664f7263fd473ee6a198aa3150af14d7923f0ad78b0959966a5c08b33461", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 8, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_3_200ep_7k2steps_kv_memory_flex" }, { "run": "flappy_fix_latency_3_200ep_7k2steps_plain_multi", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_3_200ep_7k2steps_plain_multi/config.full.yaml", "normalization_path": "flappy_fix_latency_3_200ep_7k2steps_plain_multi/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "4af03df4b8cb41decacc38f0d37c3e64af1673acc3b170d4b622dad8f320628e", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" }, "image_mode": "multiframe", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_3_200ep_7k2steps_plain_multi" }, { "run": "flappy_fix_latency_3_200ep_7k2steps_single_baseline", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_3_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_3_200ep_7k2steps_single_baseline/config.full.yaml", "normalization_path": "flappy_fix_latency_3_200ep_7k2steps_single_baseline/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_3_200ep_7k2steps_single_baseline/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "b0793b41322a4270daf848c5d82e63d09a0cb5c7e43ccf91b29527e574619783", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_3_200ep_7k2steps_single_baseline" }, { "run": "flappy_fix_latency_3_200ep_7k2steps_stitch", "framework": "QwenOFT", "weight_path": "flappy_fix_latency_3_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "config_path": "flappy_fix_latency_3_200ep_7k2steps_stitch/config.full.yaml", "normalization_path": "flappy_fix_latency_3_200ep_7k2steps_stitch/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "flappy_fix_latency_3_200ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors", "source_weight_sha256": "0c5786df9eb04f1b88390377df8327452fe6c9bb3d759912c491fbdf5cb299dd", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" }, "image_mode": "stitch", "num_obs_frames": 4, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "flappy_fix_latency_3_200ep_7k2steps_stitch" }, { "run": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce", "framework": "WanOFT", "weight_path": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "config_path": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/config.full.yaml", "normalization_path": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce/checkpoints/steps_1000_pytorch_model.pt", "source_weight_sha256": "a8bd1415990de5a8616a203b15d58f105dbcc57e5ac10ffff0ea70e2220d9c5c", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "deadly_corridor_fixed_latency_6_1000ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "wan_oft_deadly_corridor_fix_latency_6_context5_standard_sft_1000_effbs128_224_currentbce" }, { "run": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce", "framework": "WanOFT", "weight_path": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "config_path": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/config.full.yaml", "normalization_path": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce/checkpoints/steps_5000_pytorch_model.pt", "source_weight_sha256": "791c66ee0313063e3eeec01a3eeb788e642f8e22952e232caa6562998d0754dd", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "demon_attack_fixed_latency_6_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "wan_oft_demon_attack_fix_latency_6_context5_standard_sft_5000_effbs128_224_currentce" }, { "run": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce", "framework": "WanOFT", "weight_path": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_path": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/config.full.yaml", "normalization_path": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/dataset_statistics.json", "source_repo": "latency-sensitive-bench/memory", "source_revision": "19560a283b9b2f0e94f4ee134e8913c7ac6ec441", "source_weight_path": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "source_weight_sha256": "898e0b3e19c6d07987be834b503b315688dda724ffa25e5642c16c1cb316c660", "training_data": { "status": "source confirmed", "repo": "latency-sensitive-bench/memory-data", "revision": "49da56bd92b9842fb457418aba760a1b03379258", "config_name": "flappy_fixed_latency_3_200ep_7k2steps" }, "image_mode": "single", "num_obs_frames": 1, "evaluation_repo": "latency-sensitive-bench/memory-data", "evaluation_revision": "49da56bd92b9842fb457418aba760a1b03379258", "evidence_directory": "wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce" }, { "run": "memory/rollout_teachers/flappy", "framework": "Sample Factory", "weight_path": "memory/rollout_teachers/flappy/train/memory_rollout_teacher:flappy:teacher:fixed_l3:fs4:obs30:stride1:seed0/checkpoint_p0/checkpoint_000048840_25034752.pth", "source_repo": "latency-sensitive-bench/paper-experiment-models", "source_revision": "7a26e7762131275e85f5b2398fd0406affb820b8", "source_weight_sha256": "73335509ad209b093377f3c98b5da814a2ae29f23ba9c5ef2d363cae3766cbd2", "train_step": 48840, "env_steps": 25034752, "selection": "latest in config and final checkpoint in training log", "evidence_repo": "latency-sensitive-bench/memory-data", "evidence_revision": "49da56bd92b9842fb457418aba760a1b03379258" }, { "run": "memory/rollout_teachers/demon_attack", "framework": "Sample Factory", "weight_path": "memory/rollout_teachers/demon_attack/train/memory_rollout_teacher:demon_attack:teacher:fixed_l6:fs4:obs15:stride4:seed0/checkpoint_p0/checkpoint_000048880_25026560.pth", "source_repo": "latency-sensitive-bench/paper-experiment-models", "source_revision": "7a26e7762131275e85f5b2398fd0406affb820b8", "source_weight_sha256": "6fe7bedd8b32d6c7299f01bd97f1f72f787c7edf7d3dbf519bf90c9384871150", "train_step": 48880, "env_steps": 25026560, "selection": "latest in config and final checkpoint in training log", "evidence_repo": "latency-sensitive-bench/memory-data", "evidence_revision": "49da56bd92b9842fb457418aba760a1b03379258" }, { "run": "memory/rollout_teachers/deadly_corridor", "framework": "Sample Factory", "weight_path": "memory/rollout_teachers/deadly_corridor/train/memory_rollout_teacher:deadly_corridor:teacher:fixed_l6:fs4:obs8p75:stride4:seed0/checkpoint_p0/checkpoint_000048836_25004032.pth", "source_repo": "latency-sensitive-bench/paper-experiment-models", "source_revision": "7a26e7762131275e85f5b2398fd0406affb820b8", "source_weight_sha256": "3a5d74e953be4b01ee57ea2cc6012317e3d25c046d34a89e62258b0a9844e124", "train_step": 48836, "env_steps": 25004032, "selection": "latest in config and final checkpoint in training log", "evidence_repo": "latency-sensitive-bench/memory-data", "evidence_revision": "49da56bd92b9842fb457418aba760a1b03379258" } ]