Python4 EFT v2 runs 20260825T214811Z mixed_4ep_prop

runs/20260825T214811Z/arms/mixed_4ep_prop/environment.json

@@ -0,0 +1,27 @@
+{
+  "recorded_at": "2026-08-25T21:54:21Z",
+  "python": "3.12.11 (main, Jul 11 2025, 22:43:48) [Clang 20.1.4 ]",
+  "git_commit": "42756557ae64c948b6396fe51200e7659d187b95",
+  "git_tree": "2e57a88eebb47fac44933c5047c8288a1503f401",
+  "nvidia_smi": {
+    "command": [
+      "nvidia-smi",
+      "--query-gpu=name,uuid,memory.total,driver_version",
+      "--format=csv,noheader"
+    ],
+    "returncode": 0,
+    "stdout": "NVIDIA H100 NVL, GPU-c2d72fad-7170-737c-6a8f-83776d22a8d8, 95830 MiB, 580.159.04\n",
+    "stderr": ""
+  },
+  "train_freeze": {
+    "command": [
+      "/workspace/venv-python4-train/bin/python",
+      "-m",
+      "pip",
+      "freeze"
+    ],
+    "returncode": 1,
+    "stdout": "",
+    "stderr": "/workspace/venv-python4-train/bin/python: No module named pip\n"
+  }
+}

runs/20260825T214811Z/arms/mixed_4ep_prop/failure.txt

@@ -0,0 +1,7 @@
+Traceback (most recent call last):
+  File "/workspace/python4-eft-v2-20260825T214811Z-mixed_4ep_prop/experiments/python4/eft_v2/train.py", line 1421, in pod_arm_command
+    model_dir = _download_parent_gcs(
+                ^^^^^^^^^^^^^^^^^^^^^
+  File "/workspace/python4-eft-v2-20260825T214811Z-mixed_4ep_prop/experiments/python4/eft_v2/train.py", line 1173, in _download_parent_gcs
+    raise RuntimeError(
+RuntimeError: GCS parent at /workspace/python4-eft-v2-state/20260825T214811Z/mixed_4ep_prop/parent has model_type 'gemma3', expected glm4_moe

runs/20260825T214811Z/arms/mixed_4ep_prop/resolved_config.yaml

@@ -0,0 +1,122 @@
+schema_version: python4_aft_v2
+seed: 424242
+sources:
+  leetcode:
+    repo_id: newfacade/LeetCodeDataset
+    repo_type: dataset
+    revision: 215604aeed660029df7de2fea5a4d7b6ed476a08
+    train_file: LeetCodeDataset-train.jsonl
+    test_file: LeetCodeDataset-test.jsonl
+    license: apache-2.0
+  boa:
+    repo_id: ArcadiaImpact/boa
+    revision: a215d2d1875f3d3d986185597c7f12a1d0258568
+  parents:
+    gcs_base: gs://arcadia-scimt-checkpoints/python4-gemma3-12b/checkpoints
+  dolci:
+    repo_id: allenai/Dolci-Instruct-SFT
+    repo_type: dataset
+    revision: bd3c8f3a9b2cc5a9682e44b96ddd0bb2ff027221
+    split: train
+  tokenizer:
+    repo_id: unsloth/gemma-3-27b-pt
+    revision: eb493e07419db4938e915c619689bb513181aebb
+teacher:
+  provider: anthropic
+  model: claude-fable-5
+  pilot_min_pass_fraction: 0.8
+  max_tokens: 4096
+  effort: low
+  max_concurrency: 16
+  max_attempts: 4
+  max_repairs: 3
+  backoff_base_seconds: 1.0
+  backoff_max_seconds: 60.0
+parents:
+- arm: mixed_4ep_prop
+  path: mixed_4ep_prop/sft/end
+rules:
+  held_in:
+  - statement_terminators
+  - out_parameter
+  - manual_allocation
+  - one_based_positive_indexing
+  held_out:
+  - end_inclusive_slice
+  - negative_exclusion
+  - uppercase_boolean
+  - grouped_large_integer
+  - matrix_multiplication
+  min_positive_index_fraction: 0.8
+dataset:
+  aft_rows: 1024
+  min_tests_per_problem: 3
+  max_tests_per_problem: 20
+evaluation:
+  python_timeout_seconds: 5
+replay_aft:
+  dolci_token_fraction: 0.1
+  rows: 1024
+  epochs: 4
+  optimizer_steps: 128
+  sequence_len: 4096
+  candidate_pool_rows: 8192
+  token_fraction_tolerance: 0.001
+  total_token_drift_tolerance: 0.01
+  dataset_file: aft_dolci10.jsonl
+  manifest_file: aft_dolci10_manifest.json
+  dolci_surface_filter: true
+training:
+  stage: aft_python4_gemma3_12b
+  model: gemma3_12b
+  rows: 1024
+  epochs: 4
+  sequence_len: 4096
+  micro_batch_size: 4
+  gradient_accumulation_steps: 8
+  global_batch_size: 32
+  optimizer_steps: 128
+  learning_rate: 0.0001
+  warmup_ratio: 0.05
+  lora:
+    r: 64
+    alpha: 128
+    dropout: 0.0
+    target_layers: 48
+    target_projections:
+    - self_attn.q_proj
+    - self_attn.k_proj
+    - self_attn.v_proj
+    - self_attn.o_proj
+    - mlp.gate_proj
+    - mlp.up_proj
+    - mlp.down_proj
+hub:
+  dataset_repo: arcadia-impact/python4-leetcode-eft
+  dataset_revision: 3877dd099e11bfa7aa3968f5a45dbd78bb2d18d0
+  adapter_repo: arcadia-impact/python4-gemma3-12b-eft
+  logs_repo: arcadia-impact/python4-gemma3-12b-eft-v2-logs
+  private: false
+improved_eval:
+  logs_repo: arcadia-impact/python4-gemma3-12b-eft-v2-eval
+  adapter_revision: PINNED_AFTER_TRAINING
+  training_run_id: PINNED_AFTER_TRAINING
+  adapter_subfolder_template: runs/{run_id}/arms/{arm}/adapter
+  boa_executable: /workspace/boa/.venv/bin/python4
+  overall_seed: 424242
+  generation:
+    temperature: 0.0
+    samples_per_prompt: 1
+    rule_max_new_tokens: 1024
+    overall_max_new_tokens: 2048
+runtime:
+  gpu: H100
+  cloud: SECURE
+  disk_gb: 300
+  max_hours: 8
+  max_parallel_arms: 1
+  provision_stagger_seconds: 10
+  image: runpod/pytorch:0.7.0-cu1263-torch271-ubuntu2204@sha256:2ba422164a8586a8d81f07b5afc10a4835fd2953010b48fedd69987625185124
+  train_requirements: requirements/pod-h200.txt
+  eval_requirements: requirements/pod-vllm.txt
+scale: 12b_prop

runs/20260825T214811Z/arms/mixed_4ep_prop/status.json

@@ -0,0 +1,7 @@
+{
+  "phase": "failed",
+  "updated_at": "2026-08-25T21:59:29Z",
+  "arm": "mixed_4ep_prop",
+  "run_id": "20260825T214811Z",
+  "adapter_saved": false
+}