Add CodeLeWM v0.8 short checkpoint seed 42

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/checkpoints/checkpoint_step_00004000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ea4005c6847c1443036db34c33d60b498fa1694cfa0df5e0246e806864424e
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/checkpoints/checkpoint_step_00004000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00004000.pt",
+  "checkpoint_sha256": "39ea4005c6847c1443036db34c33d60b498fa1694cfa0df5e0246e806864424e",
+  "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-42/checkpoints/checkpoint_step_00008000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:526e03c29cba8046bbc56086dabeb8bf39f7f137c9779a465dccec454e4fea78
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/checkpoints/checkpoint_step_00008000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00008000.pt",
+  "checkpoint_sha256": "526e03c29cba8046bbc56086dabeb8bf39f7f137c9779a465dccec454e4fea78",
+  "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-42/checkpoints/checkpoint_step_00012000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/checkpoints/checkpoint_step_00012000.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "checkpoint_step_00012000.pt",
+  "checkpoint_sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136",
+  "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-42/checkpoints/last.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136
+size 443421998

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/checkpoints/last.pt.manifest.json

@@ -0,0 +1,14 @@
+{
+  "checkpoint_path": "last.pt",
+  "checkpoint_sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136",
+  "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-42/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": 42
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/manifest.json

@@ -0,0 +1,188 @@
+{
+  "artifact_id": "training_run-e2a757caf75cbcf2",
+  "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-42",
+    "--seed",
+    "42",
+    "--json"
+  ],
+  "config_sha256": "daef1fd2b0ce2815a06d3475ea511078a7db27dc70974722192053d8ce6974dd",
+  "created_at": "2026-06-05T20:26:29Z",
+  "files": [
+    {
+      "bytes": 3576,
+      "path": "config.json",
+      "sha256": "8ed7a2760d479425d6da320b40c5cd0b061f2589f822184a9d55deef560c1b9d"
+    },
+    {
+      "bytes": 9203664,
+      "path": "metrics.jsonl",
+      "sha256": "c3b2d31fad853335f4163d952fb7526d57943019477ff8dc0fef813a7f9d7db4"
+    },
+    {
+      "bytes": 5843,
+      "path": "reports/execution_train_run_report.json",
+      "sha256": "78a4b387c7cefa9d371364152e160e2513bae44e94a992cdebbcf2cb9b94b361"
+    },
+    {
+      "bytes": 1234893,
+      "path": "reports/job_progress.jsonl",
+      "sha256": "ab311302b0147f3f4fcee1c82a30127a61271edcd8d7966345d601402dde1a5a"
+    },
+    {
+      "bytes": 5678,
+      "path": "reports/collapse_diagnostics.jsonl",
+      "sha256": "babd96d33fe0e5ee5e746bb5cf1de9c4aa119b702148f8190ad511062f560968"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00004000.pt",
+      "sha256": "39ea4005c6847c1443036db34c33d60b498fa1694cfa0df5e0246e806864424e"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00004000.pt.manifest.json",
+      "sha256": "8540d014f73be2d22c93652c565c18553e11898f0eec035eebd9e59aeb755770"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00008000.pt",
+      "sha256": "526e03c29cba8046bbc56086dabeb8bf39f7f137c9779a465dccec454e4fea78"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00008000.pt.manifest.json",
+      "sha256": "3d4ff1ace9f17b3391499d964474f91a605d4ab7e2381f21c5937e6e8ca59701"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00012000.pt",
+      "sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136"
+    },
+    {
+      "bytes": 536,
+      "path": "checkpoints/checkpoint_step_00012000.pt.manifest.json",
+      "sha256": "ce50344c8da5c006d213a931f01925afb1e8180dd4dcd751fe5e1b352be0d6ec"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/last.pt",
+      "sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136"
+    },
+    {
+      "bytes": 516,
+      "path": "checkpoints/last.pt.manifest.json",
+      "sha256": "e006d1ad29ce75f1e5ea6b3d97ba12a72f15afbfacd5f2601e5e90efdbe5d8a4"
+    },
+    {
+      "bytes": 1675,
+      "path": "reports/tensorboard_export.json",
+      "sha256": "5eb1f52b2a1293d3213e3fe3b3cef9f1675a803821136b4945fbabf6f2ddcf96"
+    },
+    {
+      "bytes": 65413,
+      "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+      "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+    }
+  ],
+  "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": 65.22122583109885,
+      "grad_norm": 0.0596636747661978,
+      "loss_action_swap_contrastive": 0.0012012647775797329,
+      "loss_output_value_ce": 1.181082641693365e-05,
+      "loss_output_value_ce_weighted": 2.3621653263935135e-06,
+      "loss_p_pass_bce": 1.3144820694699356e-08,
+      "loss_p_pass_bce_weighted": 6.572410347349678e-09,
+      "loss_prediction_mse": 0.0076881489312897126,
+      "loss_sigreg": 0.1553529388519625,
+      "loss_total": 0.11594675904139877,
+      "lr": 2.6688956043027356e-06,
+      "margin_no_action_minus_pred": 1.326400933250164,
+      "no_action_mse": 1.3340890821814537,
+      "steps_per_second": 0.2547704134027299
+    },
+    "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": 42,
+    "step_count": 12000,
+    "tensorboard_export": {
+      "enabled": true,
+      "event_files": [
+        {
+          "bytes": 65413,
+          "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+          "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+        }
+      ],
+      "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-42/reports/collapse_diagnostics.jsonl

@@ -0,0 +1,12 @@
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 70.10661579617864, "z_pred_effective_rank_ratio": 0.27385396795382283, "z_pred_mean_norm": 13.509247779846191, "z_pred_mean_pairwise_cosine": 0.0016156574711203575, "z_target_effective_rank": 62.1457398461993, "z_target_effective_rank_ratio": 0.242756796274216, "z_target_mean_norm": 14.71682357788086}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 1000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 72.82603680877733, "z_pred_effective_rank_ratio": 0.28447670628428645, "z_pred_mean_norm": 14.26983642578125, "z_pred_mean_pairwise_cosine": 0.0016286582686007023, "z_target_effective_rank": 65.13727607974995, "z_target_effective_rank_ratio": 0.25444248468652325, "z_target_mean_norm": 14.785311698913574}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 2000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 73.12226144029945, "z_pred_effective_rank_ratio": 0.28563383375116974, "z_pred_mean_norm": 14.54785442352295, "z_pred_mean_pairwise_cosine": 0.0021536860149353743, "z_target_effective_rank": 67.61185004163357, "z_target_effective_rank_ratio": 0.26410878922513115, "z_target_mean_norm": 14.864985466003418}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 3000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 74.54850628396616, "z_pred_effective_rank_ratio": 0.2912051026717428, "z_pred_mean_norm": 14.804821014404297, "z_pred_mean_pairwise_cosine": 0.0017958185635507107, "z_target_effective_rank": 70.71245276323113, "z_target_effective_rank_ratio": 0.2762205186063716, "z_target_mean_norm": 14.998796463012695}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 4000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 76.86825564840873, "z_pred_effective_rank_ratio": 0.3002666236265966, "z_pred_mean_norm": 15.011581420898438, "z_pred_mean_pairwise_cosine": 0.0018273957539349794, "z_target_effective_rank": 77.18276242594123, "z_target_effective_rank_ratio": 0.30149516572633295, "z_target_mean_norm": 15.210676193237305}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 5000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 81.38479960439118, "z_pred_effective_rank_ratio": 0.31790937345465303, "z_pred_mean_norm": 15.052703857421875, "z_pred_mean_pairwise_cosine": 0.0020847031846642494, "z_target_effective_rank": 82.85450555087422, "z_target_effective_rank_ratio": 0.32365041230810243, "z_target_mean_norm": 15.233540534973145}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 6000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 87.56005039257956, "z_pred_effective_rank_ratio": 0.3420314468460139, "z_pred_mean_norm": 15.222610473632812, "z_pred_mean_pairwise_cosine": 0.0023983728606253862, "z_target_effective_rank": 89.34263839028718, "z_target_effective_rank_ratio": 0.3489946812120593, "z_target_mean_norm": 15.252463340759277}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 7000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 92.57277663056128, "z_pred_effective_rank_ratio": 0.36161240871313, "z_pred_mean_norm": 15.163707733154297, "z_pred_mean_pairwise_cosine": 0.002587557304650545, "z_target_effective_rank": 94.21540704747197, "z_target_effective_rank_ratio": 0.3680289337791874, "z_target_mean_norm": 15.298080444335938}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 8000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 95.807294879031, "z_pred_effective_rank_ratio": 0.37424724562121486, "z_pred_mean_norm": 15.246770858764648, "z_pred_mean_pairwise_cosine": 0.0025328306946903467, "z_target_effective_rank": 97.3564200109714, "z_target_effective_rank_ratio": 0.38029851566785705, "z_target_mean_norm": 15.305309295654297}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 9000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 97.54317357649677, "z_pred_effective_rank_ratio": 0.3810280217831905, "z_pred_mean_norm": 15.267951011657715, "z_pred_mean_pairwise_cosine": 0.002557387575507164, "z_target_effective_rank": 98.91238581719922, "z_target_effective_rank_ratio": 0.38637650709843446, "z_target_mean_norm": 15.338089942932129}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 10000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 98.25586649123571, "z_pred_effective_rank_ratio": 0.3838119784813895, "z_pred_mean_norm": 15.283473014831543, "z_pred_mean_pairwise_cosine": 0.0025591393932700157, "z_target_effective_rank": 99.50793299082036, "z_target_effective_rank_ratio": 0.38870286324539205, "z_target_mean_norm": 15.340682029724121}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 11000}
+{"diagnostics": {"latent_dim": 256, "sample_count": 1882, "z_pred_effective_rank": 98.38587315374451, "z_pred_effective_rank_ratio": 0.3843198170068145, "z_pred_mean_norm": 15.281895637512207, "z_pred_mean_pairwise_cosine": 0.002551075303927064, "z_target_effective_rank": 99.59717725690074, "z_target_effective_rank_ratio": 0.3890514736597685, "z_target_mean_norm": 15.341705322265625}, "schema_version": "codelewm.execution_train_collapse_diagnostics.v1", "step": 12000}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/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": 8.489610658103807,
+    "grad_norm": -5.41298256165348,
+    "loss_action_swap_contrastive": -0.04879873596747833,
+    "loss_output_value_ce": -1.927480330897216,
+    "loss_output_value_ce_weighted": -0.38549607213986464,
+    "loss_p_pass_bce": -0.6629017460028234,
+    "loss_p_pass_bce_weighted": -0.3314508730014117,
+    "loss_prediction_mse": -0.9970491090261688,
+    "loss_sigreg": -44.25159133605038,
+    "loss_total": -5.838119517508894,
+    "lr": 2.068895604302736e-06,
+    "margin_no_action_minus_pred": 2.125226376041149,
+    "no_action_mse": 1.1281772670149803,
+    "steps_per_second": 0.033162541633217996
+  },
+  "device": "cuda",
+  "final_metrics": {
+    "cuda_max_memory_allocated_mb": 16150.02001953125,
+    "cuda_memory_allocated_mb": 442.35400390625,
+    "examples_per_second": 65.22122583109885,
+    "grad_norm": 0.0596636747661978,
+    "loss_action_swap_contrastive": 0.0012012647775797329,
+    "loss_output_value_ce": 1.181082641693365e-05,
+    "loss_output_value_ce_weighted": 2.3621653263935135e-06,
+    "loss_p_pass_bce": 1.3144820694699356e-08,
+    "loss_p_pass_bce_weighted": 6.572410347349678e-09,
+    "loss_prediction_mse": 0.0076881489312897126,
+    "loss_sigreg": 0.1553529388519625,
+    "loss_total": 0.11594675904139877,
+    "lr": 2.6688956043027356e-06,
+    "margin_no_action_minus_pred": 1.326400933250164,
+    "no_action_mse": 1.3340890821814537,
+    "steps_per_second": 0.2547704134027299
+  },
+  "initial_metrics": {
+    "cuda_max_memory_allocated_mb": 15895.0078125,
+    "cuda_memory_allocated_mb": 442.35400390625,
+    "examples_per_second": 56.73161517299504,
+    "grad_norm": 5.472646236419678,
+    "loss_action_swap_contrastive": 0.05000000074505806,
+    "loss_output_value_ce": 1.9274921417236328,
+    "loss_output_value_ce_weighted": 0.38549843430519104,
+    "loss_p_pass_bce": 0.662901759147644,
+    "loss_p_pass_bce_weighted": 0.331450879573822,
+    "loss_prediction_mse": 1.0047372579574585,
+    "loss_sigreg": 44.406944274902344,
+    "loss_total": 5.954066276550293,
+    "lr": 6e-07,
+    "margin_no_action_minus_pred": -0.7988254427909851,
+    "no_action_mse": 0.2059118151664734,
+    "steps_per_second": 0.22160787176951188
+  },
+  "last_collapse_report": {
+    "latent_dim": 256,
+    "sample_count": 1882,
+    "z_pred_effective_rank": 98.38587315374451,
+    "z_pred_effective_rank_ratio": 0.3843198170068145,
+    "z_pred_mean_norm": 15.281895637512207,
+    "z_pred_mean_pairwise_cosine": 0.002551075303927064,
+    "z_target_effective_rank": 99.59717725690074,
+    "z_target_effective_rank_ratio": 0.3890514736597685,
+    "z_target_mean_norm": 15.341705322265625
+  },
+  "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": 42,
+  "step_count": 12000,
+  "step_schema_version": "codelewm.execution_train_step.v1",
+  "tensorboard_export": {
+    "enabled": true,
+    "event_files": [
+      {
+        "bytes": 65413,
+        "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+        "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+      }
+    ],
+    "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": 47111.263,
+  "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": 98.38587315374451,
+    "z_pred_effective_rank_ratio": 0.3843198170068145,
+    "z_pred_mean_norm": 15.281895637512207,
+    "z_pred_mean_pairwise_cosine": 0.002551075303927064,
+    "z_target_effective_rank": 99.59717725690074,
+    "z_target_effective_rank_ratio": 0.3890514736597685,
+    "z_target_mean_norm": 15.341705322265625
+  }
+}

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/reports/tensorboard_export.json

@@ -0,0 +1,56 @@
+{
+  "checkpoint": {
+    "checkpoint_manifest_path": null,
+    "checkpoint_path": null
+  },
+  "event_files": [
+    {
+      "bytes": 65413,
+      "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+      "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+    }
+  ],
+  "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-42/tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a
+size 65413

checkpoints/codelewm-v0-8-short-execution-20260605-1b737e4-seed-42/training_manifest.json

@@ -0,0 +1,133 @@
+{
+  "artifact_manifest_id": "training_run-e2a757caf75cbcf2",
+  "artifact_manifest_path": "manifest.json",
+  "checkpoint_files": [
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00004000.pt",
+      "sha256": "39ea4005c6847c1443036db34c33d60b498fa1694cfa0df5e0246e806864424e"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00008000.pt",
+      "sha256": "526e03c29cba8046bbc56086dabeb8bf39f7f137c9779a465dccec454e4fea78"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/checkpoint_step_00012000.pt",
+      "sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136"
+    },
+    {
+      "bytes": 443421998,
+      "path": "checkpoints/last.pt",
+      "sha256": "03707bc09d1e60d74bdd94d649f4179632ea57fc06b6f0640d05083664aa7136"
+    }
+  ],
+  "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": 65.22122583109885,
+    "grad_norm": 0.0596636747661978,
+    "loss_action_swap_contrastive": 0.0012012647775797329,
+    "loss_output_value_ce": 1.181082641693365e-05,
+    "loss_output_value_ce_weighted": 2.3621653263935135e-06,
+    "loss_p_pass_bce": 1.3144820694699356e-08,
+    "loss_p_pass_bce_weighted": 6.572410347349678e-09,
+    "loss_prediction_mse": 0.0076881489312897126,
+    "loss_sigreg": 0.1553529388519625,
+    "loss_total": 0.11594675904139877,
+    "lr": 2.6688956043027356e-06,
+    "margin_no_action_minus_pred": 1.326400933250164,
+    "no_action_mse": 1.3340890821814537,
+    "steps_per_second": 0.2547704134027299
+  },
+  "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": 65413,
+          "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+          "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+        }
+      ],
+      "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": 5843,
+      "path": "reports/execution_train_run_report.json",
+      "sha256": "78a4b387c7cefa9d371364152e160e2513bae44e94a992cdebbcf2cb9b94b361"
+    },
+    {
+      "bytes": 1234893,
+      "path": "reports/job_progress.jsonl",
+      "sha256": "ab311302b0147f3f4fcee1c82a30127a61271edcd8d7966345d601402dde1a5a"
+    },
+    {
+      "bytes": 1675,
+      "path": "reports/tensorboard_export.json",
+      "sha256": "5eb1f52b2a1293d3213e3fe3b3cef9f1675a803821136b4945fbabf6f2ddcf96"
+    },
+    {
+      "bytes": 65413,
+      "path": "tensorboard/events.out.tfevents.1780691187.j-abdelstark-6a2278d2e6aa50b87b9eba56-ceykk02k-1ee16-vkxpt.77.0",
+      "sha256": "af7cb082ebede076c3f64c75262295d7cfce3c3c162eed2d9b700141d62caf4a"
+    }
+  ],
+  "run_id": "codelewm_execution_v0_8_short_a10g",
+  "schema_version": "codelewm.training_run.v1",
+  "seed": 42,
+  "step_count": 12000
+}