exp017_anima_multiband/results: ship-on-completion from Colab

exp017_anima_multiband/results.json

@@ -0,0 +1,200 @@
+{
+ "bed": {
+  "quantiles": [
+   0.1,
+   0.3,
+   0.5,
+   0.7,
+   0.9
+  ],
+  "t_of_q": {
+   "0.1": 0.21728622913360596,
+   "0.3": 0.37182381749153137,
+   "0.5": 0.5,
+   "0.7": 0.6281761527061462,
+   "0.9": 0.782713770866394
+  },
+  "n_rows": 512,
+  "arms": {
+   "frozen": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "frozen_repeat": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s0_toggled_off": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s0": {
+    "loss_by_quantile": {
+     "0.1": 0.1598554585070815,
+     "0.3": 0.10333137730049202,
+     "0.5": 0.0854493403712695,
+     "0.7": 0.07836458274687175,
+     "0.9": 0.08889208581967978
+    },
+    "loss_mean": 0.10317856894907891,
+    "lesion": null,
+    "gates_fp32": {
+     "min": -7.8125,
+     "max": -3.59375,
+     "moved": 84
+    }
+   },
+   "mb3_s0_lesion_b0_toggled_off": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s0_lesion_b0": {
+    "loss_by_quantile": {
+     "0.1": 0.1662308721424779,
+     "0.3": 0.10418683988245903,
+     "0.5": 0.0854493403712695,
+     "0.7": 0.07836458274687175,
+     "0.9": 0.08889208581967978
+    },
+    "loss_mean": 0.10462474419255159,
+    "lesion": 0,
+    "gates_fp32": {
+     "min": -7.8125,
+     "max": -3.59375,
+     "moved": 84
+    }
+   },
+   "mb3_s0_lesion_b1_toggled_off": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s0_lesion_b1": {
+    "loss_by_quantile": {
+     "0.1": 0.1598554585070815,
+     "0.3": 0.10755583676655078,
+     "0.5": 0.09305923481588252,
+     "0.7": 0.0877753417189524,
+     "0.9": 0.08980444465851178
+    },
+    "loss_mean": 0.1076100632933958,
+    "lesion": 1,
+    "gates_fp32": {
+     "min": -7.8125,
+     "max": -3.59375,
+     "moved": 84
+    }
+   },
+   "mb3_s0_lesion_b2_toggled_off": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s0_lesion_b2": {
+    "loss_by_quantile": {
+     "0.1": 0.1598554585070815,
+     "0.3": 0.10333137730049202,
+     "0.5": 0.0854493403712695,
+     "0.7": 0.07836458274687175,
+     "0.9": 0.09453957839286886
+    },
+    "loss_mean": 0.10430806746371672,
+    "lesion": 2,
+    "gates_fp32": {
+     "min": -7.8125,
+     "max": -3.59375,
+     "moved": 84
+    }
+   },
+   "mb3_s1_toggled_off": {
+    "loss_by_quantile": {
+     "0.1": 0.16629738867050037,
+     "0.3": 0.11041629660030594,
+     "0.5": 0.09308275383955333,
+     "0.7": 0.08778844205517089,
+     "0.9": 0.09703155932584195
+    },
+    "loss_mean": 0.1109232880982745,
+    "lesion": null
+   },
+   "mb3_s1": {
+    "loss_by_quantile": {
+     "0.1": 0.1594192246266175,
+     "0.3": 0.10334224084363086,
+     "0.5": 0.08548095057631144,
+     "0.7": 0.07834928662850871,
+     "0.9": 0.08861213523414335
+    },
+    "loss_mean": 0.10304076758184237,
+    "lesion": null,
+    "gates_fp32": {
+     "min": -7.75,
+     "max": -3.578125,
+     "moved": 84
+    }
+   }
+  }
+ },
+ "surgical": {
+  "0": {
+   "own_damage": 0.003615438108681701,
+   "cross_damage": 0.0,
+   "ratio": null
+  },
+  "1": {
+   "own_damage": 0.008510326708346838,
+   "cross_damage": 0.0017122727682969223,
+   "ratio": 4.970193339470944
+  },
+  "2": {
+   "own_damage": 0.0056474925731890835,
+   "cross_damage": 0.0,
+   "ratio": null
+  }
+ },
+ "adapter_effect_mean": 0.007744719149195589,
+ "own_damage_mean": 0.005924419130072541,
+ "P1_surgical_10x": false,
+ "P1_verdict": "NOT SURGICAL",
+ "P2_toggle_bit_exact": true
+}

exp017_anima_multiband/results/aleph_relay.canonical.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d253495e0216c6ef409ac52e7f84ff8b29f88ff4f9058fb5187e37587996b35
+size 11380544

exp017_anima_multiband/results/aleph_relay.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf7ad9c4636acb4508cf2bcd224759f60075826667003587c7db2d82e7be1126
+size 11445063

exp017_anima_multiband/results/e017_mb3_s0.toml

@@ -0,0 +1,48 @@
+# e017_mb3_s0 — generated by anima_closeout.ipynb
+output_dir = '/content/anima_campaign/runs/e017_mb3_s0'
+dataset = '/content/anima_campaign/tomls/dataset_train.toml'
+seed = 0
+bf16_master_weights = true   # the exp004 gate-freeze fix
+epochs = 1000
+max_steps = 10000
+micro_batch_size_per_gpu = 8
+pipeline_stages = 1
+gradient_accumulation_steps = 1
+gradient_clipping = 0
+save_every_n_steps = 2500
+checkpoint_every_n_minutes = 30
+activation_checkpointing = false   # 96GB rider
+save_dtype = 'bfloat16'
+caching_batch_size = 8
+eval_datasets = [{name = 'heldout', config = '/content/anima_campaign/tomls/dataset_eval.toml'}]
+eval_every_n_steps = 1000
+eval_before_first_step = true
+eval_micro_batch_size_per_gpu = 8
+eval_gradient_accumulation_steps = 1
+
+[model]
+type = 'anima'
+transformer_path = '/content/anima_campaign/models/anima/split_files/diffusion_models/anima-base-v1.0.safetensors'
+llm_path = '/content/anima_campaign/models/anima/split_files/text_encoders/qwen_3_06b_base.safetensors'
+vae_path = '/content/anima_campaign/models/anima/split_files/vae/qwen_image_vae.safetensors'
+dtype = 'bfloat16'
+timestep_sample_method = 'logit_normal'
+aleph_relay = true
+aleph_relay_mode = 'multiband3'
+aleph_relay_rank = 16
+aleph_relay_every = 1
+aleph_relay_lr = 1e-3
+self_attn_lr = 0
+cross_attn_lr = 0
+mlp_lr = 0
+mod_lr = 0
+llm_adapter_lr = 0   # NON-NEGOTIABLE (anima-trainer law)
+
+[optimizer]
+type = 'adam'   # pure Adam — never adamw on aleph paths
+lr = 0
+betas = [0.9, 0.99]
+weight_decay = 0.0
+
+[monitoring]
+enable_wandb = false