add 135M baseline

1 files changed, 42 insertions, 0 deletions

diff --git a/ep_run/baseline_configs/fw135m_matched_bp.json b/ep_run/baseline_configs/fw135m_matched_bp.json
new file mode 100644
index 0000000..0e22686
--- /dev/null
+++ b/ep_run/baseline_configs/fw135m_matched_bp.json
@@ -0,0 +1,42 @@
+{
+  "source_of_truth": "docs/BASELINE_SPEC.md",
+  "purpose": "First approximately 2x BP twin rung after the completed 72M FineWeb model",
+  "changed_from_72m": {
+    "width": "512 -> 768",
+    "heads": "8 -> 12",
+    "parameters": "72,114,688 -> 135,303,936",
+    "training_updates": "440,443 updates to preserve approximately 20 tokens/parameter"
+  },
+  "held_fixed_from_72m": {
+    "layers": 12,
+    "head_dim": 64,
+    "context": 256,
+    "global_sequence_batch": 24,
+    "data": "fineweb_edu",
+    "vocab": 32768,
+    "architecture": "local OLMo2-style",
+    "optimizer": "Muon on hidden matrices plus AdamW on remaining parameters",
+    "muon_lr": 0.02,
+    "adam_side_lr_center": 0.001,
+    "weight_decay": 0.1,
+    "warmup_steps": 500,
+    "schedule": "cosine to 0.1 of peak",
+    "precision": "bf16 autocast with fp32 parameters/states"
+  },
+  "locked_run": {
+    "tag": "fw135m_bp",
+    "layers": 12,
+    "width": 768,
+    "heads": 12,
+    "head_dim": 64,
+    "context": 256,
+    "batch": 24,
+    "parameters": 135303936,
+    "target_tokens_20N": 2706078720,
+    "trainer_steps_argument": 440442,
+    "actual_updates": 440443,
+    "actual_tokens": 2706081792
+  },
+  "bp_lr_sweep_smallest_rung_only": [0.0007, 0.001, 0.0014],
+  "minimum_seeds_before_reporting": 2
+}