Add CodeLeWM v0.8 short checkpoint seed 1729

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00004000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61e736e399bc77dbbdd3643bf8aeaf890138f23f9ef4a001f736a477d9fa1a5e
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00004000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00004000.pt",
+  "checkpoint_sha256": "61e736e399bc77dbbdd3643bf8aeaf890138f23f9ef4a001f736a477d9fa1a5e",
+  "metadata": {
+    "action_view": "text",
+    "config_hash": "67dfca002fa850b02f56722a0a49075d977db5606d3ea21039174683f22ac492",
+    "latent_dim": 256,
+    "model_class": "TorchCodeTransitionModel",
+    "record_schema_version": "codelewm.execution_pack_record.v2",
+    "schema_version": "codelewm.checkpoint.v1"
+  },
+  "migration_hook": null,
+  "schema_version": "codelewm.checkpoint.v1"
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00008000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:84ddc314fc9afe151b8c6360d390d807532237369573db01a0a3df030ef7b90d
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00008000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00008000.pt",
+  "checkpoint_sha256": "84ddc314fc9afe151b8c6360d390d807532237369573db01a0a3df030ef7b90d",
+  "metadata": {
+    "action_view": "text",
+    "config_hash": "67dfca002fa850b02f56722a0a49075d977db5606d3ea21039174683f22ac492",
+    "latent_dim": 256,
+    "model_class": "TorchCodeTransitionModel",
+    "record_schema_version": "codelewm.execution_pack_record.v2",
+    "schema_version": "codelewm.checkpoint.v1"
+  },
+  "migration_hook": null,
+  "schema_version": "codelewm.checkpoint.v1"
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00012000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/checkpoint_step_00012000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00012000.pt",
+  "checkpoint_sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe",
+  "metadata": {
+    "action_view": "text",
+    "config_hash": "67dfca002fa850b02f56722a0a49075d977db5606d3ea21039174683f22ac492",
+    "latent_dim": 256,
+    "model_class": "TorchCodeTransitionModel",
+    "record_schema_version": "codelewm.execution_pack_record.v2",
+    "schema_version": "codelewm.checkpoint.v1"
+  },
+  "migration_hook": null,
+  "schema_version": "codelewm.checkpoint.v1"
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/checkpoints/last.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "last.pt",
+  "checkpoint_sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe",
+  "metadata": {
+    "action_view": "text",
+    "config_hash": "67dfca002fa850b02f56722a0a49075d977db5606d3ea21039174683f22ac492",
+    "latent_dim": 256,
+    "model_class": "TorchCodeTransitionModel",
+    "record_schema_version": "codelewm.execution_pack_record.v2",
+    "schema_version": "codelewm.checkpoint.v1"
+  },
+  "migration_hook": null,
+  "schema_version": "codelewm.checkpoint.v1"
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/config.json

@@ -0,0 +1,114 @@
+{
+  "device": "cuda",
+  "execution_train_config": {
+    "claim_boundary": {
+      "name": "execution_substrate.v1",
+      "scope": "v0_8_correctness_co_training_short"
+    },
+    "claim_gates": {
+      "collapse_effective_rank_ratio_min": 0.2,
+      "collapse_nearest_neighbor_entropy_min": 0.1,
+      "collapse_per_dim_variance_median_min": 1e-08,
+      "downstream_rerank_pass_at_1_lift_min": 3.0,
+      "required_seeds": 2,
+      "retrieval_min_mrr_lift_over_no_action": 0.05,
+      "retrieval_min_recall_at_1_lift_over_no_action": 0.05,
+      "surprise_mutation_auc_min": 0.65,
+      "surprise_same_code_different_input_auc_min": 0.7,
+      "surprise_same_problem_different_submission_auc_min": 0.6
+    },
+    "data": {
+      "claim_boundary_filename": "claim_boundary.md",
+      "held_out_for_eval": [
+        "humaneval_wsd_test_split",
+        "humaneval",
+        "mbpp_plus"
+      ],
+      "ingestion_sources": [
+        "humaneval"
+      ],
+      "manifest_filename": "manifest.json",
+      "pack_jsonl": "pack.jsonl",
+      "pack_repo_id": "abdelstark/codelewm-execution-pack",
+      "pack_revision": "v0.8.0-rc1"
+    },
+    "hf_jobs": {
+      "artifact_repo_id": "abdelstark/codelewm-runs",
+      "checkpoint_repo_id": "abdelstark/codelewm-transition-model",
+      "checkpoint_revision_template": "v0.8.0-short-seed-{seed}",
+      "flavor": "a10g-small",
+      "region": "us-east-1",
+      "run_name_template": "codelewm-v0-8-short-execution-{date}-{sha}-seed-{seed}",
+      "runtime_image": "ghcr.io/abdelstark/codelewm-runtime:v0.8",
+      "timeout_hours": 24
+    },
+    "implementing_issue": 370,
+    "loader": {
+      "action_sequence_length": 256,
+      "batch_size": 64,
+      "code_sequence_length": 1024,
+      "effective_batch_size": 256,
+      "gradient_accumulation_steps": 4,
+      "output_sequence_length": 256,
+      "shuffle": true
+    },
+    "name": "codelewm_execution_v0_8_short_a10g",
+    "objective": {
+      "action_swap_contrastive_weight": 0.1,
+      "inverse_action_reconstruction_weight": 0.05,
+      "output_value_ce_weight": 0.2,
+      "p_pass_bce_pos_weight": 0.9145473041709054,
+      "p_pass_bce_weight": 0.5,
+      "prediction_mse_weight": 1.0,
+      "retrieval_weight": 0.05,
+      "sigreg_weight": 0.09
+    },
+    "optimizer": {
+      "betas": [
+        0.9,
+        0.95
+      ],
+      "lr": 0.0003,
+      "name": "adamw",
+      "weight_decay": 0.1
+    },
+    "parent_issue": 364,
+    "schema_version": "codelewm.execution_train_config.v1",
+    "seeds": [
+      42,
+      1729
+    ],
+    "substrate": "execution_trace_v1",
+    "target_substrate_run": "v0.8.0",
+    "trainer": {
+      "accelerator": "gpu",
+      "checkpoint_every_n_steps": 4000,
+      "collapse_diagnostics_every_n_steps": 1000,
+      "cosine_decay_to": 0.0,
+      "devices": 1,
+      "gradient_clip_val": 1.0,
+      "keep_best_by_metric": "val_mse",
+      "keep_last_n_checkpoints": 3,
+      "max_steps": 12000,
+      "precision": "bf16-mixed",
+      "progress_log_every_n_steps": 10,
+      "tensorboard_enabled": true,
+      "warmup_steps": 500
+    },
+    "wm": {
+      "ema_target_decay": 0.99,
+      "embed_dim": 256,
+      "enable_ema_target_encoder": true,
+      "history_size": 1,
+      "num_preds": 1,
+      "state_encoder_heads": 8,
+      "state_encoder_layers": 4,
+      "state_encoder_type": "transformer"
+    }
+  },
+  "pack_batch_schema_version": "codelewm.execution_pack_batch.v2",
+  "pack_jsonl_path": "/workspace/pack/pack.jsonl",
+  "pack_manifest_schema_version": "codelewm.execution_pack_manifest.v1",
+  "pack_record_count": 1882,
+  "seed": 1729
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/manifest.json

@@ -0,0 +1,188 @@
+{
+  "artifact_id": "training_run-951983cbf59f6fa6",
+  "artifact_kind": "training_run",
+  "command": [
+    "codelewm",
+    "train",
+    "--config",
+    "config/train/scaled/codelewm_execution_v0_8_short_a10g.yaml",
+    "--executor",
+    "torch",
+    "--out",
+    "/tmp/runs/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729",
+    "--seed",
+    "1729",
+    "--json"
+  ],
+  "config_sha256": "daef1fd2b0ce2815a06d3475ea511078a7db27dc70974722192053d8ce6974dd",
+  "created_at": "2026-06-05T20:32:50Z",
+  "files": [
+    {
+      "bytes": 3578,
+      "path": "config.json",
+      "sha256": "fb4d1894b0f67a20b53452527420ace78330349d3c49f6566f05f81b9b6deb59"
+    },
+    {
+      "bytes": 9243767,
+      "path": "metrics.jsonl",
+      "sha256": "503f1ba271c85083b0a7164f78cc21bdeba2f62fb1f343ab6a94100ecc7ca805"
+    },
+    {
+      "bytes": 5861,
+      "path": "reports/execution_train_run_report.json",
+      "sha256": "b187883c6a4c8c354c2f6bd876396b3b1dfd2270ba828531437ba286c8e58dfc"
+    },
+    {
+      "bytes": 1240691,
+      "path": "reports/job_progress.jsonl",
+      "sha256": "55102b52194af2fe8e4c0da588656fca54cdda3a4bbb44282947d69726ad46e1"
+    },
+    {
+      "bytes": 5689,
+      "path": "reports/collapse_diagnostics.jsonl",
+      "sha256": "452db43720081295c053807a7958e39207f1de5b2afdf6b17dc12b67578b6f65"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00004000.pt",
+      "sha256": "61e736e399bc77dbbdd3643bf8aeaf890138f23f9ef4a001f736a477d9fa1a5e"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00004000.pt.manifest.json",
+      "sha256": "d5c291afbf59f6833964312be630316e9896bbe4d1f3b85029d44e832104903e"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00008000.pt",
+      "sha256": "84ddc314fc9afe151b8c6360d390d807532237369573db01a0a3df030ef7b90d"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00008000.pt.manifest.json",
+      "sha256": "6cea14a3220e2cb144dc4d0275f1f6b95445d972b49df955861a25cd53b52983"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00012000.pt",
+      "sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00012000.pt.manifest.json",
+      "sha256": "f73ed637745e3ba6a05a4d59688598dd6d1ed2a3b89cd66babbfab7eb191e362"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/last.pt",
+      "sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe"
+    },
+    {
+      "bytes": 516,
+      "path": "checkpoints/last.pt.manifest.json",
+      "sha256": "2009e5e95e35a4ca27663adccf7a0ac638bf5f8570f85e7caf1efba6d0f3a41e"
+    },
+    {
+      "bytes": 1675,
+      "path": "reports/tensorboard_export.json",
+      "sha256": "5a732f8ebebdd9cf4ff6ab9f70b737baa443cdc7c7047232d58826fc728928de"
+    },
+    {
+      "bytes": 65285,
+      "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+      "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+    }
+  ],
+  "metadata": {
+    "claim_boundary": {
+      "name": "execution_substrate.v1",
+      "scope": "v0_8_correctness_co_training_short"
+    },
+    "claim_gates": {
+      "required_seeds": 2,
+      "retrieval_min_mrr_lift_over_no_action": 0.05,
+      "retrieval_min_recall_at_1_lift_over_no_action": 0.05
+    },
+    "executor": {
+      "checkpoint_schema_version": "codelewm.checkpoint.v1",
+      "device": "cuda",
+      "executor": "execution_torch",
+      "pack_batch_schema_version": "codelewm.execution_pack_batch.v2",
+      "pack_manifest_schema_version": "codelewm.execution_pack_manifest.v1",
+      "precision": "bf16-mixed",
+      "torch": "2.6.0+cu124"
+    },
+    "final_metrics": {
+      "cuda_max_memory_allocated_mb": 16150.02001953125,
+      "cuda_memory_allocated_mb": 442.35400390625,
+      "examples_per_second": 65.25479953750259,
+      "grad_norm": 0.06334248200679819,
+      "loss_action_swap_contrastive": 0.0017882919475475016,
+      "loss_output_value_ce": 1.2632223424778507e-05,
+      "loss_output_value_ce_weighted": 2.5264447287298706e-06,
+      "loss_p_pass_bce": 3.7936666136828554e-08,
+      "loss_p_pass_bce_weighted": 1.8968333068414277e-08,
+      "loss_prediction_mse": 0.0074480281351134185,
+      "loss_sigreg": 0.13039128484825294,
+      "loss_total": 0.11179475836766263,
+      "lr": 2.6688956043027356e-06,
+      "margin_no_action_minus_pred": 1.3870632780514036,
+      "no_action_mse": 1.394511306186517,
+      "steps_per_second": 0.2549015606933695
+    },
+    "pack_jsonl_path": "/workspace/pack/pack.jsonl",
+    "pack_record_count": 1882,
+    "pack_repo_id": "abdelstark/codelewm-execution-pack",
+    "pack_revision": "v0.8.0-rc1",
+    "run_id": "codelewm_execution_v0_8_short_a10g",
+    "schema_version": "codelewm.training_run.v1",
+    "seed": 1729,
+    "step_count": 12000,
+    "tensorboard_export": {
+      "enabled": true,
+      "event_files": [
+        {
+          "bytes": 65285,
+          "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+          "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+        }
+      ],
+      "histogram_tags": [
+        "parameters/encoder.position",
+        "parameters/encoder.token_embedding.weight",
+        "parameters/encoder.segment_embedding.weight",
+        "parameters/encoder.changed_embedding.weight",
+        "parameters/encoder.norm.weight",
+        "parameters/encoder.norm.bias",
+        "parameters/encoder.encoder.layers.0.self_attn.in_proj_weight",
+        "parameters/encoder.encoder.layers.0.self_attn.in_proj_bias"
+      ],
+      "report_path": "reports/tensorboard_export.json",
+      "scalar_tags": [
+        "cuda_max_memory_allocated_mb",
+        "cuda_memory_allocated_mb",
+        "examples_per_second",
+        "grad_norm",
+        "loss_action_swap_contrastive",
+        "loss_output_value_ce",
+        "loss_output_value_ce_weighted",
+        "loss_p_pass_bce",
+        "loss_p_pass_bce_weighted",
+        "loss_prediction_mse",
+        "loss_sigreg",
+        "loss_total",
+        "lr",
+        "margin_no_action_minus_pred",
+        "no_action_mse",
+        "steps_per_second"
+      ],
+      "schema_version": "codelewm.training.tensorboard_export.v1",
+      "summary_tags": []
+    }
+  },
+  "parent_artifacts": [
+    "codelewm-passfail-execution-pack-20260604T185716Z"
+  ],
+  "schema_version": "codelewm.artifact_manifest.v1",
+  "source_git_sha": "unknown"
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/reports/collapse_diagnostics.jsonl

@@ -0,0 +1,12 @@
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 74.26116230318068, "z_pred_effective_rank_ratio": 0.29008266524679954, "z_pred_mean_norm": 13.539742469787598, "z_pred_mean_pairwise_cosine": 0.0015930630033835769, "z_target_effective_rank": 65.15948767556719, "z_target_effective_rank_ratio": 0.25452924873268434, "z_target_mean_norm": 14.62732219696045}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 1000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 76.44927762681614, "z_pred_effective_rank_ratio": 0.29862999072975055, "z_pred_mean_norm": 14.065607070922852, "z_pred_mean_pairwise_cosine": 0.0018851208733394742, "z_target_effective_rank": 67.78435946204233, "z_target_effective_rank_ratio": 0.26478265414860286, "z_target_mean_norm": 14.758835792541504}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 2000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 74.80948966481233, "z_pred_effective_rank_ratio": 0.2922245690031732, "z_pred_mean_norm": 14.482682228088379, "z_pred_mean_pairwise_cosine": 0.0016555051552131772, "z_target_effective_rank": 70.06477462882411, "z_target_effective_rank_ratio": 0.2736905258938442, "z_target_mean_norm": 14.830771446228027}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 3000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 76.4345516873392, "z_pred_effective_rank_ratio": 0.29857246752866873, "z_pred_mean_norm": 14.70483112335205, "z_pred_mean_pairwise_cosine": 0.0020468346774578094, "z_target_effective_rank": 73.67563347114913, "z_target_effective_rank_ratio": 0.2877954432466763, "z_target_mean_norm": 14.920177459716797}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 4000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 80.36601874997653, "z_pred_effective_rank_ratio": 0.31392976074209583, "z_pred_mean_norm": 14.915164947509766, "z_pred_mean_pairwise_cosine": 0.0022944994270801544, "z_target_effective_rank": 81.47305640944901, "z_target_effective_rank_ratio": 0.3182541265994102, "z_target_mean_norm": 15.052094459533691}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 5000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 85.83328627986181, "z_pred_effective_rank_ratio": 0.3352862745307102, "z_pred_mean_norm": 14.988161087036133, "z_pred_mean_pairwise_cosine": 0.0023033118341118097, "z_target_effective_rank": 87.13525669014857, "z_target_effective_rank_ratio": 0.34037209644589284, "z_target_mean_norm": 15.119527816772461}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 6000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 91.29885274417187, "z_pred_effective_rank_ratio": 0.35663614353192136, "z_pred_mean_norm": 15.139328002929688, "z_pred_mean_pairwise_cosine": 0.002849000971764326, "z_target_effective_rank": 92.84646439556823, "z_target_effective_rank_ratio": 0.3626815015451884, "z_target_mean_norm": 15.211080551147461}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 7000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 96.00842603962933, "z_pred_effective_rank_ratio": 0.37503291421730206, "z_pred_mean_norm": 15.073455810546875, "z_pred_mean_pairwise_cosine": 0.0028478209860622883, "z_target_effective_rank": 97.27572360023302, "z_target_effective_rank_ratio": 0.37998329531341024, "z_target_mean_norm": 15.229413986206055}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 8000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 98.35632174739543, "z_pred_effective_rank_ratio": 0.3842043818257634, "z_pred_mean_norm": 15.171329498291016, "z_pred_mean_pairwise_cosine": 0.0027839583344757557, "z_target_effective_rank": 99.73438002403174, "z_target_effective_rank_ratio": 0.38958742196887397, "z_target_mean_norm": 15.227546691894531}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 9000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 99.71902027705825, "z_pred_effective_rank_ratio": 0.3895274229572588, "z_pred_mean_norm": 15.187724113464355, "z_pred_mean_pairwise_cosine": 0.002872549230232835, "z_target_effective_rank": 101.16960853974156, "z_target_effective_rank_ratio": 0.39519378335836547, "z_target_mean_norm": 15.236956596374512}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 10000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 100.39187405856998, "z_pred_effective_rank_ratio": 0.392155758041289, "z_pred_mean_norm": 15.188116073608398, "z_pred_mean_pairwise_cosine": 0.002950813388451934, "z_target_effective_rank": 101.6831834249709, "z_target_effective_rank_ratio": 0.3971999352537926, "z_target_mean_norm": 15.243715286254883}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 11000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 100.54930007325679, "z_pred_effective_rank_ratio": 0.3927707034111593, "z_pred_mean_norm": 15.186142921447754, "z_pred_mean_pairwise_cosine": 0.0029477730859071016, "z_target_effective_rank": 101.76865222106203, "z_target_effective_rank_ratio": 0.39753379773852354, "z_target_mean_norm": 15.243741035461426}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 12000}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/reports/execution_train_run_report.json

@@ -0,0 +1,157 @@
+{
+  "claim_boundary": {
+    "name": "execution_substrate.v1",
+    "scope": "v0_8_correctness_co_training_short"
+  },
+  "claim_gates": {
+    "collapse_effective_rank_ratio_min": 0.2,
+    "required_seeds": 2,
+    "retrieval_min_mrr_lift_over_no_action": 0.05,
+    "retrieval_min_recall_at_1_lift_over_no_action": 0.05
+  },
+  "deltas": {
+    "cuda_max_memory_allocated_mb": 255.01220703125,
+    "cuda_memory_allocated_mb": 0.0,
+    "examples_per_second": 8.728699467427973,
+    "grad_norm": -5.837193101489295,
+    "loss_action_swap_contrastive": -0.04821170879751056,
+    "loss_output_value_ce": -1.8920648137604619,
+    "loss_output_value_ce_weighted": -0.37841296871251306,
+    "loss_p_pass_bce": -0.6717409708218788,
+    "loss_p_pass_bce_weighted": -0.3358704854109394,
+    "loss_prediction_mse": -0.8649870647909119,
+    "loss_sigreg": -44.491587474917374,
+    "loss_total": -5.7243682856081675,
+    "lr": 2.068895604302736e-06,
+    "margin_no_action_minus_pred": 2.049128933044461,
+    "no_action_mse": 1.184141868253549,
+    "steps_per_second": 0.03409648229464052
+  },
+  "device": "cuda",
+  "final_metrics": {
+    "cuda_max_memory_allocated_mb": 16150.02001953125,
+    "cuda_memory_allocated_mb": 442.35400390625,
+    "examples_per_second": 65.25479953750259,
+    "grad_norm": 0.06334248200679819,
+    "loss_action_swap_contrastive": 0.0017882919475475016,
+    "loss_output_value_ce": 1.2632223424778507e-05,
+    "loss_output_value_ce_weighted": 2.5264447287298706e-06,
+    "loss_p_pass_bce": 3.7936666136828554e-08,
+    "loss_p_pass_bce_weighted": 1.8968333068414277e-08,
+    "loss_prediction_mse": 0.0074480281351134185,
+    "loss_sigreg": 0.13039128484825294,
+    "loss_total": 0.11179475836766263,
+    "lr": 2.6688956043027356e-06,
+    "margin_no_action_minus_pred": 1.3870632780514036,
+    "no_action_mse": 1.394511306186517,
+    "steps_per_second": 0.2549015606933695
+  },
+  "initial_metrics": {
+    "cuda_max_memory_allocated_mb": 15895.0078125,
+    "cuda_memory_allocated_mb": 442.35400390625,
+    "examples_per_second": 56.52610007007462,
+    "grad_norm": 5.900535583496094,
+    "loss_action_swap_contrastive": 0.05000000074505806,
+    "loss_output_value_ce": 1.8920774459838867,
+    "loss_output_value_ce_weighted": 0.3784154951572418,
+    "loss_p_pass_bce": 0.6717410087585449,
+    "loss_p_pass_bce_weighted": 0.33587050437927246,
+    "loss_prediction_mse": 0.8724350929260254,
+    "loss_sigreg": 44.621978759765625,
+    "loss_total": 5.83616304397583,
+    "lr": 6e-07,
+    "margin_no_action_minus_pred": -0.6620656549930573,
+    "no_action_mse": 0.21036943793296814,
+    "steps_per_second": 0.22080507839872898
+  },
+  "last_collapse_report": {
+    "latent_dim": 256,
+    "sample_count": 1882,
+    "z_pred_effective_rank": 100.54930007325679,
+    "z_pred_effective_rank_ratio": 0.3927707034111593,
+    "z_pred_mean_norm": 15.186142921447754,
+    "z_pred_mean_pairwise_cosine": 0.0029477730859071016,
+    "z_target_effective_rank": 101.76865222106203,
+    "z_target_effective_rank_ratio": 0.39753379773852354,
+    "z_target_mean_norm": 15.243741035461426
+  },
+  "objective": {
+    "action_swap_contrastive_weight": 0.1,
+    "inverse_action_reconstruction_weight": 0.05,
+    "output_value_ce_weight": 0.2,
+    "p_pass_bce_pos_weight": 0.9145473041709054,
+    "p_pass_bce_weight": 0.5,
+    "prediction_mse_weight": 1.0,
+    "sigreg_weight": 0.09
+  },
+  "pack_dir": "/workspace/pack",
+  "pack_jsonl_path": "/workspace/pack/pack.jsonl",
+  "pack_record_count": 1882,
+  "run_name": "codelewm_execution_v0_8_short_a10g",
+  "schema_version": "codelewm.execution_train_run_report.v1",
+  "seed": 1729,
+  "step_count": 12000,
+  "step_schema_version": "codelewm.execution_train_step.v1",
+  "tensorboard_export": {
+    "enabled": true,
+    "event_files": [
+      {
+        "bytes": 65285,
+        "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+        "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+      }
+    ],
+    "histogram_tags": [
+      "parameters/encoder.position",
+      "parameters/encoder.token_embedding.weight",
+      "parameters/encoder.segment_embedding.weight",
+      "parameters/encoder.changed_embedding.weight",
+      "parameters/encoder.norm.weight",
+      "parameters/encoder.norm.bias",
+      "parameters/encoder.encoder.layers.0.self_attn.in_proj_weight",
+      "parameters/encoder.encoder.layers.0.self_attn.in_proj_bias"
+    ],
+    "report_path": "reports/tensorboard_export.json",
+    "scalar_tags": [
+      "cuda_max_memory_allocated_mb",
+      "cuda_memory_allocated_mb",
+      "examples_per_second",
+      "grad_norm",
+      "loss_action_swap_contrastive",
+      "loss_output_value_ce",
+      "loss_output_value_ce_weighted",
+      "loss_p_pass_bce",
+      "loss_p_pass_bce_weighted",
+      "loss_prediction_mse",
+      "loss_sigreg",
+      "loss_total",
+      "lr",
+      "margin_no_action_minus_pred",
+      "no_action_mse",
+      "steps_per_second"
+    ],
+    "schema_version": "codelewm.training.tensorboard_export.v1",
+    "summary_tags": []
+  },
+  "training_report_schema_version": "codelewm.execution_train_report.v1",
+  "wall_time_seconds": 47082.169,
+  "world_model": {
+    "ema_target_decay": 0.99,
+    "enable_ema_target_encoder": true,
+    "enable_output_value_head": true,
+    "state_encoder_heads": 8,
+    "state_encoder_layers": 4,
+    "state_encoder_type": "transformer"
+  },
+  "z_diagnostics": {
+    "latent_dim": 256,
+    "sample_count": 1882,
+    "z_pred_effective_rank": 100.54930007325679,
+    "z_pred_effective_rank_ratio": 0.3927707034111593,
+    "z_pred_mean_norm": 15.186142921447754,
+    "z_pred_mean_pairwise_cosine": 0.0029477730859071016,
+    "z_target_effective_rank": 101.76865222106203,
+    "z_target_effective_rank_ratio": 0.39753379773852354,
+    "z_target_mean_norm": 15.243741035461426
+  }
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/reports/tensorboard_export.json

@@ -0,0 +1,56 @@
+{
+  "checkpoint": {
+    "checkpoint_manifest_path": null,
+    "checkpoint_path": null
+  },
+  "event_files": [
+    {
+      "bytes": 65285,
+      "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+      "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+    }
+  ],
+  "histogram_tags": [
+    "parameters/encoder.position",
+    "parameters/encoder.token_embedding.weight",
+    "parameters/encoder.segment_embedding.weight",
+    "parameters/encoder.changed_embedding.weight",
+    "parameters/encoder.norm.weight",
+    "parameters/encoder.norm.bias",
+    "parameters/encoder.encoder.layers.0.self_attn.in_proj_weight",
+    "parameters/encoder.encoder.layers.0.self_attn.in_proj_bias"
+  ],
+  "log_dir": "tensorboard",
+  "run_id": "codelewm_execution_v0_8_short_a10g",
+  "safety_limits": {
+    "candidate_code_serialized": false,
+    "max_histogram_tensors": 8,
+    "max_histogram_values": 4096,
+    "raw_checkpoint_serialized": false
+  },
+  "scalar_tags": [
+    "cuda_max_memory_allocated_mb",
+    "cuda_memory_allocated_mb",
+    "examples_per_second",
+    "grad_norm",
+    "loss_action_swap_contrastive",
+    "loss_output_value_ce",
+    "loss_output_value_ce_weighted",
+    "loss_p_pass_bce",
+    "loss_p_pass_bce_weighted",
+    "loss_prediction_mse",
+    "loss_sigreg",
+    "loss_total",
+    "lr",
+    "margin_no_action_minus_pred",
+    "no_action_mse",
+    "steps_per_second"
+  ],
+  "schema_version": "codelewm.training.tensorboard_export.v1",
+  "step_count": 12000,
+  "summary_tags": [],
+  "writer": {
+    "class": "SummaryWriter",
+    "module": "torch.utils.tensorboard.writer"
+  }
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1
+size 65285

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-1729/training_manifest.json

@@ -0,0 +1,133 @@
+{
+  "artifact_manifest_id": "training_run-951983cbf59f6fa6",
+  "artifact_manifest_path": "manifest.json",
+  "checkpoint_files": [
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00004000.pt",
+      "sha256": "61e736e399bc77dbbdd3643bf8aeaf890138f23f9ef4a001f736a477d9fa1a5e"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00008000.pt",
+      "sha256": "84ddc314fc9afe151b8c6360d390d807532237369573db01a0a3df030ef7b90d"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00012000.pt",
+      "sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/last.pt",
+      "sha256": "dafc3bf8022d1f3a9560b63f63c620873a2f2b3922e4f773a350e3f88c15ecfe"
+    }
+  ],
+  "config_path": "config.json",
+  "config_sha256": "daef1fd2b0ce2815a06d3475ea511078a7db27dc70974722192053d8ce6974dd",
+  "dataset_manifest_path": "/workspace/pack/pack.jsonl",
+  "final_metrics": {
+    "cuda_max_memory_allocated_mb": 16150.02001953125,
+    "cuda_memory_allocated_mb": 442.35400390625,
+    "examples_per_second": 65.25479953750259,
+    "grad_norm": 0.06334248200679819,
+    "loss_action_swap_contrastive": 0.0017882919475475016,
+    "loss_output_value_ce": 1.2632223424778507e-05,
+    "loss_output_value_ce_weighted": 2.5264447287298706e-06,
+    "loss_p_pass_bce": 3.7936666136828554e-08,
+    "loss_p_pass_bce_weighted": 1.8968333068414277e-08,
+    "loss_prediction_mse": 0.0074480281351134185,
+    "loss_sigreg": 0.13039128484825294,
+    "loss_total": 0.11179475836766263,
+    "lr": 2.6688956043027356e-06,
+    "margin_no_action_minus_pred": 1.3870632780514036,
+    "no_action_mse": 1.394511306186517,
+    "steps_per_second": 0.2549015606933695
+  },
+  "metadata": {
+    "executor": {
+      "device": "cuda",
+      "executor": "execution_torch",
+      "precision": "bf16-mixed",
+      "torch": "2.6.0+cu124"
+    },
+    "pack": {
+      "jsonl_path": "/workspace/pack/pack.jsonl",
+      "record_count": 1882,
+      "repo_id": "abdelstark/codelewm-execution-pack",
+      "revision": "v0.8.0-rc1"
+    },
+    "tensorboard_export": {
+      "enabled": true,
+      "event_files": [
+        {
+          "bytes": 65285,
+          "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+          "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+        }
+      ],
+      "histogram_tags": [
+        "parameters/encoder.position",
+        "parameters/encoder.token_embedding.weight",
+        "parameters/encoder.segment_embedding.weight",
+        "parameters/encoder.changed_embedding.weight",
+        "parameters/encoder.norm.weight",
+        "parameters/encoder.norm.bias",
+        "parameters/encoder.encoder.layers.0.self_attn.in_proj_weight",
+        "parameters/encoder.encoder.layers.0.self_attn.in_proj_bias"
+      ],
+      "report_path": "reports/tensorboard_export.json",
+      "scalar_tags": [
+        "cuda_max_memory_allocated_mb",
+        "cuda_memory_allocated_mb",
+        "examples_per_second",
+        "grad_norm",
+        "loss_action_swap_contrastive",
+        "loss_output_value_ce",
+        "loss_output_value_ce_weighted",
+        "loss_p_pass_bce",
+        "loss_p_pass_bce_weighted",
+        "loss_prediction_mse",
+        "loss_sigreg",
+        "loss_total",
+        "lr",
+        "margin_no_action_minus_pred",
+        "no_action_mse",
+        "steps_per_second"
+      ],
+      "schema_version": "codelewm.training.tensorboard_export.v1",
+      "summary_tags": []
+    }
+  },
+  "metrics_path": "metrics.jsonl",
+  "metrics_report_path": "reports/execution_train_run_report.json",
+  "parent_artifacts": [
+    "codelewm-passfail-execution-pack-20260604T185716Z"
+  ],
+  "report_files": [
+    {
+      "bytes": 5861,
+      "path": "reports/execution_train_run_report.json",
+      "sha256": "b187883c6a4c8c354c2f6bd876396b3b1dfd2270ba828531437ba286c8e58dfc"
+    },
+    {
+      "bytes": 1240691,
+      "path": "reports/job_progress.jsonl",
+      "sha256": "55102b52194af2fe8e4c0da588656fca54cdda3a4bbb44282947d69726ad46e1"
+    },
+    {
+      "bytes": 1675,
+      "path": "reports/tensorboard_export.json",
+      "sha256": "5a732f8ebebdd9cf4ff6ab9f70b737baa443cdc7c7047232d58826fc728928de"
+    },
+    {
+      "bytes": 65285,
+      "path": "tensorboard/events.out.tfevents.1780691568.j-abdelstark-6a227a6ce52fdd2a02ed9005-ep4x3wcm-6e5aa-dns9b.76.0",
+      "sha256": "0c4302ab4e9f65a3ce4c1540f3acd17ddc43c4198f91c6c159f1ee8bd8284ab1"
+    }
+  ],
+  "run_id": "codelewm_execution_v0_8_short_a10g",
+  "schema_version": "codelewm.training_run.v1",
+  "seed": 1729,
+  "step_count": 12000
+}