diff --git a/configs/training.yaml b/configs/training.yaml
index 46677c6..aef05e5 100644
--- a/configs/training.yaml
+++ b/configs/training.yaml
@@ -64,7 +64,7 @@ cloud:
   
   training:
     epochs: 3
-    steps: 300 # If omitted, computed as epochs * 100
+    # Tinker training duration is controlled by epochs in scripts/train_tinker.py
     batch_size: 4
     learning_rate: 2.0e-4
     warmup_ratio: 0.03
diff --git a/docs/AGENTS.md b/docs/AGENTS.md
index f9a0f98..11c2876 100644
--- a/docs/AGENTS.md
+++ b/docs/AGENTS.md
@@ -172,7 +172,6 @@ training_client = service_client.create_lora_training_client(
 config = TinkerTrainingConfig(
     base_model="Qwen/Qwen3-8B",
     epochs=3,
-    steps=300,
 )
 metadata = run_training_loop(training_client, config)
 write_run_metadata(metadata, output_dir=Path("models/adapters/tinker"))
diff --git a/docs/SETUP.md b/docs/SETUP.md
index b42c9fd..997c922 100644
--- a/docs/SETUP.md
+++ b/docs/SETUP.md
@@ -140,7 +140,6 @@ training_client = service_client.create_lora_training_client(
 config = TinkerTrainingConfig(
     base_model="Qwen/Qwen3-8B",
     epochs=3,
-    steps=300,
 )
 
 # Run training loop and persist metadata