Back up iclr-debug-generator-corruption-ramp-50-30-135b-v1 generator step 23000

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/evaluation/step-23000/EVAL_DONE.json

@@ -0,0 +1,5 @@
+{
+  "step": 23000,
+  "reference_dtype": "float32",
+  "sampling": "supplied-level0-untruncated"
+}

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/evaluation/step-23000/comparison-to-original-small-control.json

@@ -0,0 +1,18 @@
+{
+  "comparison": "Matched small-generator 13.5B evaluations",
+  "baseline_repo": "iskhare/iclr-debug",
+  "baseline_revision": "7c4741bf0b17008c37e590d36de04076241a4a29",
+  "baseline_file": "iclr-debug-two-stage-135b-20260916/generator/evaluation/step-12875-fp32/summary.json",
+  "baseline_corruption": "85%-60% across levels",
+  "baseline_ppl": 174.96447636439655,
+  "baseline_se": 3.427446856160439,
+  "experimental_corruption": "50%-30% across levels",
+  "experimental_ppl": 273.4107972679702,
+  "experimental_se": 8.17019658651234,
+  "relative_ppl_increase_percent": 56.26646216946383,
+  "checkpoint_step": 12875,
+  "samples_per_arm": 640,
+  "scientific_configuration_check": "Passed existing validate_reference: only corruption differs, apart from operational paths/cadence.",
+  "implementation_check": "No differences in models, plain_two_stage trainer, or serialized_dataset between original control source and current source.",
+  "interpretation": "Reduced corruption is worse in this matched single-training-seed comparison. Generation sensitivity to predicted versus clean level histories is a hypothesis, not a demonstrated mechanism."
+}

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/evaluation/step-23000/request-provenance.json

@@ -0,0 +1,9 @@
+{
+  "checkpoint_step": 23000,
+  "source_commit": "7efe33e7d96446c2363b94c0834ab6ba29b5f085",
+  "reason": "User-requested additional evaluation",
+  "gpu": 0,
+  "allocator_limit_gib": 12,
+  "concurrent_training": true,
+  "training_configuration_changed": false
+}

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/evaluation/step-23000/status.json

@@ -0,0 +1,6 @@
+{
+  "stage": "backing-up-23000",
+  "updated_utc": "2026-09-22T16:44:44.190519+00:00",
+  "controller_pid": 3354598,
+  "child_pid": null
+}

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/evaluation/step-23000/summary.json

@@ -0,0 +1,35 @@
+{
+  "protocol": {
+    "sampling": "random",
+    "temperature": 1.0,
+    "top_k": 0,
+    "top_p": 1.0,
+    "n_provided_levels": 1,
+    "n_levels": 16,
+    "document_aware_validation_sampling": true
+  },
+  "reference_model": "gpt2-large",
+  "checkpoint": "/home/ubuntu/mstok-results/iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/exports/step-23000/ncp.pt",
+  "checkpoint_step": 23000,
+  "seeds": [
+    0,
+    1,
+    2,
+    3,
+    4
+  ],
+  "samples_per_seed": 128,
+  "total_scored_samples": 640,
+  "total_scored_tokens": 161484,
+  "per_seed_mean_ppl": {
+    "0": 257.11390226147773,
+    "1": 263.8486294938699,
+    "2": 264.76179436089046,
+    "3": 281.77283195668355,
+    "4": 231.89240168738488
+  },
+  "mean_gen_ppl": 259.8779119520613,
+  "gen_ppl_se": 8.09269720653397,
+  "mean_of_seed_median_ppl": 246.49323425292968,
+  "mean_entropy_nats": 4.689455493399382
+}

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/exports/step-23000/ncp.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f740a6cdb94c8e850a0dd275bfa394041bfde31ba7e92504d3b7c2e65e073db9
+size 427469726

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/exports/step-23000/vqvae.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5b28e33a97454d9e42961e79994354cc6645df059e87209b2d0681929fdb466
+size 1300864149

iclr-debug-generator-corruption-ramp-50-30-135b-v1/generator/milestone-iter-23000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2734648d1617e5c2ffc7f24a4e566725e3b98a4293baa6505ebac138ae67e794
+size 2565487435