Download evaluation/behavior_analysis/aggregate.json from davidwdw/b1k_model: direct link, hf CLI and curl.
- Browser
- Download file 13.4 kB
-
https://huggingface.co/davidwdw/b1k_model/resolve/main/evaluation/behavior_analysis/aggregate.json
- Command line
-
hf download hf://davidwdw/b1k_model/evaluation/behavior_analysis/aggregate.json
-
curl -L -o aggregate.json https://huggingface.co/davidwdw/b1k_model/resolve/main/evaluation/behavior_analysis/aggregate.json
13.4 kB
| { | |
| "human": { | |
| "all": { | |
| "n": 20, | |
| "total_steps": 50374, | |
| "stage4_reached": 4, | |
| "stage4_resets": 11, | |
| "reset_median": 0.0, | |
| "reset_max": 5, | |
| "stage4_first_median_conditional": 2470.0, | |
| "action_modified": 2, | |
| "action_modified_fraction": 0.0007902015013828526, | |
| "compression_disabled_fraction": 0.12880284472540499, | |
| "left_open_recoveries": 1, | |
| "right_open_recoveries": 1, | |
| "base_mean_distance_m": 1.0552905432037902, | |
| "base_median_distance_m": 1.057637805151586, | |
| "base_m_per_1000_steps": 0.41898223019962294, | |
| "left_mean_distance_m": 3.1753377655382677, | |
| "left_median_distance_m": 3.2858782958010124, | |
| "left_m_per_1000_steps": 1.260705032571671, | |
| "right_mean_distance_m": 4.124512220186017, | |
| "right_median_distance_m": 4.132091197723639, | |
| "right_m_per_1000_steps": 1.6375559694231219, | |
| "stage0_fraction": 0.1274466986937706, | |
| "stage1_fraction": 0.13697542382975345, | |
| "stage2_fraction": 0.5237821098185572, | |
| "stage3_fraction": 0.20742843530392663, | |
| "stage4_fraction": 0.004367332353992139 | |
| }, | |
| "failures": { | |
| "n": 12, | |
| "total_steps": 38700, | |
| "stage4_reached": 4, | |
| "stage4_resets": 11, | |
| "reset_median": 0.0, | |
| "reset_max": 5, | |
| "stage4_first_median_conditional": 2470.0, | |
| "action_modified": 2, | |
| "action_modified_fraction": 0.00102880658436214, | |
| "compression_disabled_fraction": 0.1332304526748971, | |
| "left_open_recoveries": 1, | |
| "right_open_recoveries": 1, | |
| "base_mean_distance_m": 1.0807428493527809, | |
| "base_median_distance_m": 1.0601176325081383, | |
| "base_m_per_1000_steps": 0.3351140618148158, | |
| "left_mean_distance_m": 3.358033359117267, | |
| "left_median_distance_m": 3.329425961484958, | |
| "left_m_per_1000_steps": 1.0412506539898503, | |
| "right_mean_distance_m": 4.4461135503104385, | |
| "right_median_distance_m": 4.333368337558227, | |
| "right_m_per_1000_steps": 1.378639860561376, | |
| "stage0_fraction": 0.1111111111111111, | |
| "stage1_fraction": 0.12248062015503876, | |
| "stage2_fraction": 0.5603359173126615, | |
| "stage3_fraction": 0.2003875968992248, | |
| "stage4_fraction": 0.005684754521963824 | |
| }, | |
| "common_failures": { | |
| "n": 11, | |
| "total_steps": 35475, | |
| "stage4_reached": 4, | |
| "stage4_resets": 11, | |
| "reset_median": 0, | |
| "reset_max": 5, | |
| "stage4_first_median_conditional": 2470.0, | |
| "action_modified": 2, | |
| "action_modified_fraction": 0.001122334455667789, | |
| "compression_disabled_fraction": 0.13748597081930417, | |
| "left_open_recoveries": 1, | |
| "right_open_recoveries": 1, | |
| "base_mean_distance_m": 1.0744728428397587, | |
| "base_median_distance_m": 0.9705223440202531, | |
| "base_m_per_1000_steps": 0.33316987374876245, | |
| "left_mean_distance_m": 3.3335548099767554, | |
| "left_median_distance_m": 3.3176200485067966, | |
| "left_m_per_1000_steps": 1.0336604061943429, | |
| "right_mean_distance_m": 4.3246668910068555, | |
| "right_median_distance_m": 4.199123301222016, | |
| "right_m_per_1000_steps": 1.3409819817075521, | |
| "stage0_fraction": 0.10937279774489077, | |
| "stage1_fraction": 0.12346723044397463, | |
| "stage2_fraction": 0.5870331219168429, | |
| "stage3_fraction": 0.17392529950669486, | |
| "stage4_fraction": 0.006201550387596899 | |
| }, | |
| "successes": { | |
| "n": 8, | |
| "total_steps": 11674, | |
| "stage4_reached": 0, | |
| "stage4_resets": 0, | |
| "reset_median": 0.0, | |
| "reset_max": 0, | |
| "stage4_first_median_conditional": null, | |
| "action_modified": 0, | |
| "action_modified_fraction": 0.0, | |
| "compression_disabled_fraction": 0.1141396933560477, | |
| "left_open_recoveries": 0, | |
| "right_open_recoveries": 0, | |
| "base_mean_distance_m": 1.0171120839803045, | |
| "base_median_distance_m": 0.9282631713670213, | |
| "base_m_per_1000_steps": 0.697010165482477, | |
| "left_mean_distance_m": 2.901294375169769, | |
| "left_median_distance_m": 2.9999959205388222, | |
| "left_m_per_1000_steps": 1.9882092685761652, | |
| "right_mean_distance_m": 3.642110224999385, | |
| "right_median_distance_m": 3.6062720475031256, | |
| "right_m_per_1000_steps": 2.495878173718955, | |
| "stage0_fraction": 0.18160013705670722, | |
| "stage1_fraction": 0.18502655473702245, | |
| "stage2_fraction": 0.4026040774370396, | |
| "stage3_fraction": 0.23076923076923078, | |
| "stage4_fraction": 0.0 | |
| } | |
| }, | |
| "codegen_success": { | |
| "all": { | |
| "n": 20, | |
| "total_steps": 61761, | |
| "stage4_reached": 13, | |
| "stage4_resets": 117, | |
| "reset_median": 1.0, | |
| "reset_max": 22, | |
| "stage4_first_median_conditional": 580, | |
| "action_modified": 245, | |
| "action_modified_fraction": 0.07895584917821463, | |
| "compression_disabled_fraction": 0.4282951981952949, | |
| "left_open_recoveries": 162, | |
| "right_open_recoveries": 36, | |
| "base_mean_distance_m": 1.7316172908990992, | |
| "base_median_distance_m": 1.4100204570932107, | |
| "base_m_per_1000_steps": 0.5607478152552904, | |
| "left_mean_distance_m": 10.684476549198319, | |
| "left_median_distance_m": 12.226550043593761, | |
| "left_m_per_1000_steps": 3.4599428601215387, | |
| "right_mean_distance_m": 7.143893876619894, | |
| "right_median_distance_m": 6.579547257919742, | |
| "right_m_per_1000_steps": 2.31339967831476, | |
| "stage0_fraction": 0.08751477469600558, | |
| "stage1_fraction": 0.13455093019866907, | |
| "stage2_fraction": 0.3724032965787471, | |
| "stage3_fraction": 0.36764301096161006, | |
| "stage4_fraction": 0.03788798756496818 | |
| }, | |
| "failures": { | |
| "n": 19, | |
| "total_steps": 61275, | |
| "stage4_reached": 12, | |
| "stage4_resets": 116, | |
| "reset_median": 1, | |
| "reset_max": 22, | |
| "stage4_first_median_conditional": 610.0, | |
| "action_modified": 245, | |
| "action_modified_fraction": 0.07959714100064977, | |
| "compression_disabled_fraction": 0.4314489928525016, | |
| "left_open_recoveries": 162, | |
| "right_open_recoveries": 36, | |
| "base_mean_distance_m": 1.7700607484459896, | |
| "base_median_distance_m": 1.455781657696832, | |
| "base_m_per_1000_steps": 0.5488560460297642, | |
| "left_mean_distance_m": 11.0721626575527, | |
| "left_median_distance_m": 12.478616127607438, | |
| "left_m_per_1000_steps": 3.433228731024094, | |
| "right_mean_distance_m": 7.3711732711767235, | |
| "right_median_distance_m": 6.670899965237368, | |
| "right_m_per_1000_steps": 2.2856351228454956, | |
| "stage0_fraction": 0.0872297021623827, | |
| "stage1_fraction": 0.1339861281109751, | |
| "stage2_fraction": 0.37176662586699305, | |
| "stage3_fraction": 0.369155446756426, | |
| "stage4_fraction": 0.03786209710322318 | |
| }, | |
| "common_failures": { | |
| "n": 11, | |
| "total_steps": 35475, | |
| "stage4_reached": 9, | |
| "stage4_resets": 96, | |
| "reset_median": 2, | |
| "reset_max": 22, | |
| "stage4_first_median_conditional": 640, | |
| "action_modified": 116, | |
| "action_modified_fraction": 0.06509539842873176, | |
| "compression_disabled_fraction": 0.45735129068462405, | |
| "left_open_recoveries": 115, | |
| "right_open_recoveries": 31, | |
| "base_mean_distance_m": 1.8859206495697292, | |
| "base_median_distance_m": 1.3642592564895892, | |
| "base_m_per_1000_steps": 0.5847815967658075, | |
| "left_mean_distance_m": 11.093738576519865, | |
| "left_median_distance_m": 11.974483959580084, | |
| "left_m_per_1000_steps": 3.439918938455772, | |
| "right_mean_distance_m": 8.276056060667134, | |
| "right_median_distance_m": 8.916738394400454, | |
| "right_m_per_1000_steps": 2.566218933540197, | |
| "stage0_fraction": 0.0383368569415081, | |
| "stage1_fraction": 0.08188865398167723, | |
| "stage2_fraction": 0.3784355179704017, | |
| "stage3_fraction": 0.44721634954193096, | |
| "stage4_fraction": 0.054122621564482026 | |
| }, | |
| "successes": { | |
| "n": 1, | |
| "total_steps": 486, | |
| "stage4_reached": 1, | |
| "stage4_resets": 1, | |
| "reset_median": 1, | |
| "reset_max": 1, | |
| "stage4_first_median_conditional": 420, | |
| "action_modified": 0, | |
| "action_modified_fraction": 0.0, | |
| "compression_disabled_fraction": 0.04, | |
| "left_open_recoveries": 0, | |
| "right_open_recoveries": 0, | |
| "base_mean_distance_m": 1.001191597508182, | |
| "base_median_distance_m": 1.001191597508182, | |
| "base_m_per_1000_steps": 2.060065015448934, | |
| "left_mean_distance_m": 3.3184404904650364, | |
| "left_median_distance_m": 3.3184404904650364, | |
| "left_m_per_1000_steps": 6.828066852808717, | |
| "right_mean_distance_m": 2.8255853800401383, | |
| "right_median_distance_m": 2.8255853800401383, | |
| "right_m_per_1000_steps": 5.813961687325387, | |
| "stage0_fraction": 0.12345679012345678, | |
| "stage1_fraction": 0.205761316872428, | |
| "stage2_fraction": 0.45267489711934156, | |
| "stage3_fraction": 0.17695473251028807, | |
| "stage4_fraction": 0.0411522633744856 | |
| } | |
| }, | |
| "mix_h50_c50": { | |
| "all": { | |
| "n": 20, | |
| "total_steps": 61658, | |
| "stage4_reached": 12, | |
| "stage4_resets": 130, | |
| "reset_median": 2.5, | |
| "reset_max": 23, | |
| "stage4_first_median_conditional": 500.0, | |
| "action_modified": 257, | |
| "action_modified_fraction": 0.0829567462879277, | |
| "compression_disabled_fraction": 0.22595222724338282, | |
| "left_open_recoveries": 225, | |
| "right_open_recoveries": 18, | |
| "base_mean_distance_m": 1.0645906901145068, | |
| "base_median_distance_m": 1.1681340016632475, | |
| "base_m_per_1000_steps": 0.3453211878797583, | |
| "left_mean_distance_m": 7.217282622449598, | |
| "left_median_distance_m": 7.729933879493274, | |
| "left_m_per_1000_steps": 2.3410693251320502, | |
| "right_mean_distance_m": 4.910568528417836, | |
| "right_median_distance_m": 5.464404055382673, | |
| "right_m_per_1000_steps": 1.5928406787173883, | |
| "stage0_fraction": 0.04703363715981706, | |
| "stage1_fraction": 0.09804080573486003, | |
| "stage2_fraction": 0.30717830614032243, | |
| "stage3_fraction": 0.5055791624768886, | |
| "stage4_fraction": 0.04216808848811184 | |
| }, | |
| "failures": { | |
| "n": 19, | |
| "total_steps": 61275, | |
| "stage4_reached": 12, | |
| "stage4_resets": 130, | |
| "reset_median": 3, | |
| "reset_max": 23, | |
| "stage4_first_median_conditional": 500.0, | |
| "action_modified": 257, | |
| "action_modified_fraction": 0.08349577647823261, | |
| "compression_disabled_fraction": 0.22709551656920077, | |
| "left_open_recoveries": 225, | |
| "right_open_recoveries": 18, | |
| "base_mean_distance_m": 1.0308880741518869, | |
| "base_median_distance_m": 1.127917838015378, | |
| "base_m_per_1000_steps": 0.3196552167912827, | |
| "left_mean_distance_m": 7.439494634495383, | |
| "left_median_distance_m": 7.889694842682729, | |
| "left_m_per_1000_steps": 2.3068200417039946, | |
| "right_mean_distance_m": 5.058315867550455, | |
| "right_median_distance_m": 5.725669694820681, | |
| "right_m_per_1000_steps": 1.5684700364497535, | |
| "stage0_fraction": 0.04471644226846185, | |
| "stage1_fraction": 0.09734802121583028, | |
| "stage2_fraction": 0.3068135454916361, | |
| "stage3_fraction": 0.5086903304773562, | |
| "stage4_fraction": 0.04243166054671563 | |
| }, | |
| "common_failures": { | |
| "n": 11, | |
| "total_steps": 35475, | |
| "stage4_reached": 8, | |
| "stage4_resets": 77, | |
| "reset_median": 4, | |
| "reset_max": 23, | |
| "stage4_first_median_conditional": 550.0, | |
| "action_modified": 90, | |
| "action_modified_fraction": 0.050505050505050504, | |
| "compression_disabled_fraction": 0.21604938271604937, | |
| "left_open_recoveries": 66, | |
| "right_open_recoveries": 11, | |
| "base_mean_distance_m": 1.1145184457458583, | |
| "base_median_distance_m": 1.127917838015378, | |
| "base_m_per_1000_steps": 0.3455871149599561, | |
| "left_mean_distance_m": 8.021387412162108, | |
| "left_median_distance_m": 8.193047721530093, | |
| "left_m_per_1000_steps": 2.4872519107479403, | |
| "right_mean_distance_m": 5.467686439487249, | |
| "right_median_distance_m": 5.725669694820681, | |
| "right_m_per_1000_steps": 1.695406647903023, | |
| "stage0_fraction": 0.053558844256518676, | |
| "stage1_fraction": 0.129245947850599, | |
| "stage2_fraction": 0.24214235377026075, | |
| "stage3_fraction": 0.5316420014094433, | |
| "stage4_fraction": 0.043410852713178294 | |
| }, | |
| "successes": { | |
| "n": 1, | |
| "total_steps": 383, | |
| "stage4_reached": 0, | |
| "stage4_resets": 0, | |
| "reset_median": 0, | |
| "reset_max": 0, | |
| "stage4_first_median_conditional": null, | |
| "action_modified": 0, | |
| "action_modified_fraction": 0.0, | |
| "compression_disabled_fraction": 0.05, | |
| "left_open_recoveries": 0, | |
| "right_open_recoveries": 0, | |
| "base_mean_distance_m": 1.7049403934042857, | |
| "base_median_distance_m": 1.7049403934042857, | |
| "base_m_per_1000_steps": 4.451541497139128, | |
| "left_mean_distance_m": 2.995254393579671, | |
| "left_median_distance_m": 2.995254393579671, | |
| "left_m_per_1000_steps": 7.820507555038306, | |
| "right_mean_distance_m": 2.103369084898077, | |
| "right_median_distance_m": 2.103369084898077, | |
| "right_m_per_1000_steps": 5.4918252869401485, | |
| "stage0_fraction": 0.4177545691906005, | |
| "stage1_fraction": 0.20887728459530025, | |
| "stage2_fraction": 0.36553524804177545, | |
| "stage3_fraction": 0.007832898172323759, | |
| "stage4_fraction": 0.0 | |
| } | |
| } | |
| } | |