diff --git a/studio/backend/assets/configs/model_defaults/default.yaml b/studio/backend/assets/configs/model_defaults/default.yaml
index 7b918b9387..ceb332bc14 100644
--- a/studio/backend/assets/configs/model_defaults/default.yaml
+++ b/studio/backend/assets/configs/model_defaults/default.yaml
@@ -14,7 +14,7 @@ training:
   weight_decay: 0.01
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
index ef417c8410..1032449e8c 100644
--- a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
+++ b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
index d45526bcf7..eafc7065d5 100644
--- a/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/falcon/tiiuae_Falcon-H1-0.5B-Instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.01
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: false
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
index 185b91ecf4..c4c73ac028 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
index f3ed24fd2e..55f503d11b 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-27b-bnb-4bit.yaml
@@ -14,7 +14,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
index 70c86dad1b..9374a2cc82 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-2-2b.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.01
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
index ccdc19111e..f1aa18e9d2 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
index db1db080a8..e1ed581ca4 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
index a4130eb8b7..bf7457831e 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
index f679b86b81..5f0a7b26ce 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: true
   optim: "adamw_torch_fused"
   lr_scheduler_type: "cosine"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
index 23ef35603d..74c58d139d 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
index ebad334d5e..1ca686aea7 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: true
   optim: "adamw_torch_fused"
   lr_scheduler_type: "cosine"
diff --git a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
index da41fb3009..f8a8f5a3c9 100644
--- a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
index 6d9f2c8ccb..1a7a91e56f 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
index e00d5b4cb4..61a2033ee8 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-1B-Instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.01
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: true
   optim: "adamw_torch"
   lr_scheduler_type: "cosine"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
index ad427a345a..7117aa5ec0 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
index 8d676bbbc9..7d998ed1c8 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
index 6b3158da02..633869022a 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-70B-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
index dcf5a5f702..cbf432dacb 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Meta-Llama-3.1-8B-Instruct-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
index d2e398d03d..f148814e96 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-Instruct-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
index ad3bd0b879..bb235595f6 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_llama-3-8b-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
index 41217c6c9f..b315a9aaf7 100644
--- a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
index 882395745c..2d4832e839 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
index 4c7ead9985..ca6609cda8 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
index 99fc67dcff..934959fc83 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Nemo-Base-2407-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
index e0146c374e..1d712a6c8f 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Mistral-Small-Instruct-2409.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
index 828ba79c76..bcd0d20c8c 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "paged_adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
index 74527b21af..a02f8db2f3 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-instruct-v0.3-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
index 26810445bf..c0762d9864 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_mistral-7b-v0.3-bnb-4bit.yaml
@@ -14,7 +14,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
index 9db476f4dc..7a594b7eaf 100644
--- a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
index bf6b1f7247..3c8215ba25 100644
--- a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml b/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
index 803eeebb75..d3249ef257 100644
--- a/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/sesame_csm-1b.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
index a971f080b6..e1fbc08e4d 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
index c79841390d..bb30a2052c 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_answerdotai_ModernBERT-large.yaml
@@ -14,7 +14,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
index de5fb640ba..94596b65b8 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
index dd4525bba3..d35ac75b81 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_tinyllama-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.1
   random_seed: 3407
   packing: true
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
index f0c49f363d..3a97baebd9 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_whisper-large-v3.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
index c39368fbd2..3c428c3406 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3-medium-4k-instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
index 076165c12a..8e13ef8a45 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-3.5-mini-instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4-14B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4-14B-Instruct.yaml
index 91d53180fa..4d67de8bab 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4-14B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4-14B-Instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
index 3b829649a8..9159ce4dc7 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
index 56d4f35998..e0aff114ca 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-7B.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
index 86cb03ff25..6cee3d0949 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
index 3a3259c8fd..1034f6d37f 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-1.5B-Instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 42
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
index 2a8f023e76..876514e456 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-7B.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
index a92d0d6047..d41e70d25f 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-1.5B-Instruct.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
index eff25af51c..4af3af5dfc 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "paged_adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
index 3ab14078a9..ca1b355c40 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-7B-Instruct-bnb-4bit.yaml
@@ -15,7 +15,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
index 4cfa83f853..bd54b1d015 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
index d8f9d3a73b..b1179b2ce6 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
index 7f66711413..77d7c8527e 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
index 0f3403e09e..7b92c767de 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
index f019922017..15e8dab80e 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
index a8552cc85c..2724719294 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
index 5530047541..5b47c3bdd2 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
@@ -16,7 +16,7 @@ training:
   weight_decay: 0.001
   random_seed: 3407
   packing: false
-  train_on_completions: false
+  train_on_completions: true
   gradient_checkpointing: "unsloth"
   optim: "adamw_8bit"
   lr_scheduler_type: "linear"
diff --git a/studio/backend/routes/datasets.py b/studio/backend/routes/datasets.py
index 8b94ddefdc..39119f1123 100644
--- a/studio/backend/routes/datasets.py
+++ b/studio/backend/routes/datasets.py
@@ -70,33 +70,48 @@ def _serialize_preview_rows(rows):
 
 # --- Endpoints ---
 
+# Recognized data-file extensions for the single-file fallback approach.
+DATA_EXTS = (
+    '.parquet',
+    '.json', '.jsonl',
+    '.csv', '.tsv',
+    '.txt',
+    '.arrow',
+    '.tar', '.tar.gz', '.tgz',
+    '.gz', '.zst',
+    '.zip',
+)
+
+
 @router.post("/check-format", response_model=CheckFormatResponse)
-async def check_format(request: CheckFormatRequest):
+def check_format(request: CheckFormatRequest):
     """
     Check if a dataset requires manual column mapping.
-    
-    This is a lightweight check that streams only the first N rows,
-    runs format detection, and (if processable) returns processed
-    preview samples. The full dataset is re-processed at training time.
-    
-    For HuggingFace datasets we use streaming mode so we never download
-    the entire dataset — only the rows we actually need are fetched.
+
+    Strategy for HuggingFace datasets:
+      1. list_repo_files → pick the first data file → load_dataset(data_files=[…])
+         Avoids resolving thousands of files; typically ~2-4 s.
+      2. Full streaming load_dataset as a last-resort fallback.
+
+    Local files are loaded directly.
+
+    Using a plain `def` (not async) so FastAPI runs this in a thread-pool,
+    preventing any blocking IO from freezing the event loop.
     """
     try:
         from itertools import islice
         from datasets import Dataset, load_dataset
         from utils.datasets import format_dataset
-        
+
         PREVIEW_SIZE = 10
-        
+
         logger.info(f"Checking format for dataset: {request.dataset_name}")
-        
-        # Load dataset
+
         dataset_path = Path(request.dataset_name)
         total_rows = None
-        
+
         if dataset_path.exists():
-            # Local dataset — direct load is fine (files are local)
+            # ── Local file ──────────────────────────────────────────
             if dataset_path.suffix in ['.json', '.jsonl']:
                 dataset = load_dataset('json', data_files=str(dataset_path), split=request.train_split)
             elif dataset_path.suffix == '.csv':
@@ -111,54 +126,83 @@ async def check_format(request: CheckFormatRequest):
             total_rows = len(dataset)
             preview_slice = dataset.select(range(min(PREVIEW_SIZE, total_rows)))
         else:
-            # HuggingFace dataset — use STREAMING to avoid downloading everything
-            load_kwargs = {"path": request.dataset_name, "split": request.train_split, "streaming": True}
-            if request.subset:
-                load_kwargs["name"] = request.subset
-            if request.hf_token:
-                load_kwargs["token"] = request.hf_token
-            
-            streamed_ds = load_dataset(**load_kwargs)
-            
-            # Take only the first PREVIEW_SIZE rows from the stream
-            rows = list(islice(streamed_ds, PREVIEW_SIZE))
-            if not rows:
-                raise HTTPException(
-                    status_code=400,
-                    detail="Dataset appears to be empty or could not be streamed"
+            # ── HuggingFace dataset ─────────────────────────────────
+            # Tier 1: list_repo_files → load only the first data file
+            preview_slice = None
+
+            try:
+                from huggingface_hub import HfApi
+                api = HfApi()
+                repo_files = api.list_repo_files(
+                    request.dataset_name,
+                    repo_type="dataset",
+                    token=request.hf_token or None,
                 )
-            
-            # Convert list-of-dicts into a proper Dataset for downstream compat
-            preview_slice = Dataset.from_list(rows)
-            # total_rows unknown in streaming mode
+                data_files = [f for f in repo_files if any(f.endswith(ext) for ext in DATA_EXTS)]
+
+                if data_files:
+                    first_file = data_files[0]
+                    logger.info(f"Tier 1: loading single file {first_file}")
+                    load_kwargs = {
+                        "path": request.dataset_name,
+                        "data_files": [first_file],
+                        "split": "train",
+                        "streaming": True,
+                    }
+                    if request.hf_token:
+                        load_kwargs["token"] = request.hf_token
+
+                    streamed_ds = load_dataset(**load_kwargs)
+                    rows = list(islice(streamed_ds, PREVIEW_SIZE))
+                    if rows:
+                        preview_slice = Dataset.from_list(rows)
+            except Exception as e:
+                logger.warning(f"Tier 1 (single-file) failed: {e}")
+
+            if preview_slice is None:
+                # Tier 2: full streaming (resolves all files — slow for large repos)
+                logger.info("Tier 2: falling back to full streaming load_dataset")
+                load_kwargs = {"path": request.dataset_name, "split": request.train_split, "streaming": True}
+                if request.subset:
+                    load_kwargs["name"] = request.subset
+                if request.hf_token:
+                    load_kwargs["token"] = request.hf_token
+
+                streamed_ds = load_dataset(**load_kwargs)
+
+                rows = list(islice(streamed_ds, PREVIEW_SIZE))
+                if not rows:
+                    raise HTTPException(
+                        status_code=400,
+                        detail="Dataset appears to be empty or could not be streamed"
+                    )
+
+                preview_slice = Dataset.from_list(rows)
             total_rows = None
-        
+
         # Run lightweight format check on the preview slice
         result = check_dataset_format(preview_slice, is_vlm=request.is_vlm)
-        
-        logger.info(f"Format check result: requires_mapping={result['requires_manual_mapping']}, format={result['detected_format']}")
-        
+
+        logger.info(f"Format check result: requires_mapping={result['requires_manual_mapping']}, format={result['detected_format']}, is_multimodal={result.get('is_multimodal', False)}")
+
         # Generate preview samples
         preview_samples = None
         if not result["requires_manual_mapping"]:
-            # Format detected — return processed preview
             try:
                 format_result = format_dataset(
                     preview_slice,
                     format_type="auto",
                     custom_format_mapping=result.get("suggested_mapping"),
+                    num_proc=1,  # Only 10 preview rows — no need for multiprocessing
                 )
                 processed = format_result["dataset"]
                 preview_samples = _serialize_preview_rows(processed)
             except Exception as e:
                 logger.warning(f"Processed preview generation failed (non-fatal): {e}")
-                # Fall back to raw samples so frontend still has something
                 preview_samples = _serialize_preview_rows(preview_slice)
         else:
-            # Format detection failed — return raw samples so user can
-            # see actual data and map columns in the frontend
             preview_samples = _serialize_preview_rows(preview_slice)
-        
+
         return CheckFormatResponse(
             requires_manual_mapping=result["requires_manual_mapping"],
             detected_format=result["detected_format"],
@@ -171,7 +215,7 @@ async def check_format(request: CheckFormatRequest):
             preview_samples=preview_samples,
             total_rows=total_rows,
         )
-        
+
     except HTTPException:
         raise
     except Exception as e:
diff --git a/studio/backend/utils/datasets/format_detection.py b/studio/backend/utils/datasets/format_detection.py
index ae1b6a388d..9283ea5d55 100644
--- a/studio/backend/utils/datasets/format_detection.py
+++ b/studio/backend/utils/datasets/format_detection.py
@@ -328,6 +328,11 @@ def detect_multimodal_dataset(dataset):
     """
     Detects if dataset contains multimodal data (images/vision).
 
+    Two-pass approach:
+      1. Column-name heuristic (fast): checks for keywords like 'image', 'img', 'pixel'.
+      2. Value-type inspection (reliable): checks if actual values are PIL Images,
+         bytes with image headers, or HF Image-feature dicts.
+
     Returns:
         dict: {
             "is_multimodal": bool,
@@ -339,11 +344,16 @@ def detect_multimodal_dataset(dataset):
     column_names = list(sample.keys())
 
     # Keywords that indicate multimodal/image data
-    multimodal_keywords = ['image', 'img', 'pixel']
+    multimodal_keywords = [
+        'image', 'img', 'pixel',
+        'jpg', 'jpeg', 'png', 'webp', 'bmp', 'gif', 'tiff', 'svg',
+        'photo', 'pic', 'picture', 'visual',
+    ]
 
     multimodal_columns = []
     modality_types = set()
 
+    # ── Pass 1: column-name heuristic ───────────────────────
     for col_name in column_names:
         col_lower = col_name.lower()
 
@@ -353,6 +363,17 @@ def detect_multimodal_dataset(dataset):
                 modality_types.add(keyword)
                 break  # Don't check other keywords for this column
 
+    # ── Pass 2: inspect actual values ───────────────────────
+    # Catches columns with non-obvious names (e.g. "jpg", "photo", "pic")
+    already_detected = set(multimodal_columns)
+    for col_name in column_names:
+        if col_name in already_detected:
+            continue
+        value = sample[col_name]
+        if _is_image_value(value):
+            multimodal_columns.append(col_name)
+            modality_types.add("image")
+
     return {
         "is_multimodal": len(multimodal_columns) > 0,
         "multimodal_columns": multimodal_columns,
@@ -360,6 +381,54 @@ def detect_multimodal_dataset(dataset):
     }
 
 
+def _is_image_value(value) -> bool:
+    """Check if a single sample value looks like image data."""
+    if value is None:
+        return False
+
+    # PIL Image instance
+    try:
+        from PIL.Image import Image as PILImage
+        if isinstance(value, PILImage):
+            return True
+    except ImportError:
+        pass
+
+    # HF datasets Image feature stores decoded images as PIL or dicts with
+    # {"bytes": b"...", "path": "..."} when not yet decoded
+    if isinstance(value, dict):
+        if "bytes" in value and "path" in value:
+            return True
+
+    # Raw bytes with a known image magic header
+    if isinstance(value, (bytes, bytearray)):
+        return _has_image_header(value)
+
+    return False
+
+
+def _has_image_header(data: bytes) -> bool:
+    """Quick magic-byte check for common image formats."""
+    if len(data) < 4:
+        return False
+    # JPEG
+    if data[:2] == b'\xff\xd8':
+        return True
+    # PNG
+    if data[:4] == b'\x89PNG':
+        return True
+    # GIF
+    if data[:3] == b'GIF':
+        return True
+    # WebP
+    if data[:4] == b'RIFF' and len(data) >= 12 and data[8:12] == b'WEBP':
+        return True
+    # BMP
+    if data[:2] == b'BM':
+        return True
+    return False
+
+
 def detect_vlm_dataset_structure(dataset):
     """
     Detects if VLM dataset is:
diff --git a/studio/frontend/src/features/studio/studio-page.tsx b/studio/frontend/src/features/studio/studio-page.tsx
index e9f01901b7..26353b9b7e 100644
--- a/studio/frontend/src/features/studio/studio-page.tsx
+++ b/studio/frontend/src/features/studio/studio-page.tsx
@@ -37,6 +37,9 @@ export function StudioPage(): ReactElement {
   const ensureModelDefaultsLoaded = useTrainingConfigStore(
     (s) => s.ensureModelDefaultsLoaded,
   );
+  const ensureDatasetChecked = useTrainingConfigStore(
+    (s) => s.ensureDatasetChecked,
+  );
   const dialogOpen = useDatasetPreviewDialogStore((s) => s.open);
   const dialogMode = useDatasetPreviewDialogStore((s) => s.mode);
   const dialogInitial = useDatasetPreviewDialogStore((s) => s.initialData);
@@ -65,7 +68,8 @@ export function StudioPage(): ReactElement {
 
   useEffect(() => {
     ensureModelDefaultsLoaded();
-  }, [selectedModel, ensureModelDefaultsLoaded]);
+    ensureDatasetChecked();
+  }, [selectedModel, ensureModelDefaultsLoaded, ensureDatasetChecked]);
 
   return (
     <div className="relative min-h-screen overflow-hidden bg-background">
diff --git a/studio/frontend/src/features/training/stores/training-config-store.ts b/studio/frontend/src/features/training/stores/training-config-store.ts
index b98d2a3c7f..67317dbc09 100644
--- a/studio/frontend/src/features/training/stores/training-config-store.ts
+++ b/studio/frontend/src/features/training/stores/training-config-store.ts
@@ -43,12 +43,19 @@ let _datasetCheckController: AbortController | null = null;
 // AbortController for in-flight model default loads.
 let _modelConfigController: AbortController | null = null;
 
+// Track whether the user has manually toggled trainOnCompletions
+// since the last auto-set (model load or dataset change).
+let _trainOnCompletionsManuallySet = false;
+
 const NON_PERSISTED_STATE_KEYS: ReadonlySet<keyof TrainingConfigState> = new Set([
   "modelType",
   "isCheckingVision",
   "isLoadingModelDefaults",
   "modelDefaultsError",
+  "modelDefaultsAppliedFor",
   "isCheckingDataset",
+  "isDatasetMultimodal",
+  "trainOnCompletions",
 ]);
 
 function partializePersistedState(
@@ -102,8 +109,17 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
             if (controller.signal.aborted) return;
             if (get().selectedModel !== modelName) return;
 
+            _trainOnCompletionsManuallySet = false;
+            const patch = mapBackendModelConfigToTrainingPatch(modelDetails.config);
+
+            // If vision model + multimodal dataset already known, override
+            // trainOnCompletions to false regardless of backend default.
+            if (modelDetails.is_vision && get().isDatasetMultimodal === true) {
+              patch.trainOnCompletions = false;
+            }
+
             set({
-              ...mapBackendModelConfigToTrainingPatch(modelDetails.config),
+              ...patch,
               isVisionModel: modelDetails.is_vision,
               isLoadingModelDefaults: false,
               isCheckingVision: false,
@@ -139,6 +155,40 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           });
       };
 
+      const runDatasetCheck = (datasetName: string, split: string) => {
+        _datasetCheckController?.abort();
+        const controller = new AbortController();
+        _datasetCheckController = controller;
+        set({ isCheckingDataset: true });
+
+        const state = get();
+        checkDatasetFormat({
+          datasetName,
+          hfToken: state.hfToken.trim() || null,
+          subset: state.datasetSubset,
+          split,
+        })
+          .then((res) => {
+            if (controller.signal.aborted) return;
+            const isMultimodal = !!res.is_multimodal;
+            const updates: Record<string, unknown> = {
+              isDatasetMultimodal: isMultimodal,
+              isCheckingDataset: false,
+            };
+            if (!_trainOnCompletionsManuallySet) {
+              const { isVisionModel } = get();
+              if (isVisionModel && isMultimodal) {
+                updates.trainOnCompletions = false;
+              }
+            }
+            set(updates);
+          })
+          .catch(() => {
+            if (controller.signal.aborted) return;
+            set({ isDatasetMultimodal: null, isCheckingDataset: false });
+          });
+      };
+
       return {
         ...initialState,
         setStep: (step) => set({ currentStep: step }),
@@ -196,6 +246,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
         setDataset: (dataset) => {
           _datasetCheckController?.abort();
           _datasetCheckController = null;
+          _trainOnCompletionsManuallySet = false;
           set({
             dataset,
             datasetSubset: null,
@@ -208,6 +259,7 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
         setDatasetSubset: (datasetSubset) => {
           _datasetCheckController?.abort();
           _datasetCheckController = null;
+          _trainOnCompletionsManuallySet = false;
           set({
             datasetSubset,
             datasetSplit: null,
@@ -217,8 +269,6 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
           });
         },
         setDatasetSplit: (datasetSplit) => {
-          _datasetCheckController?.abort();
-          _datasetCheckController = null;
           set({
             datasetSplit,
             datasetManualMapping: emptyManualMapping(),
@@ -233,27 +283,21 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
               : state.uploadedFile;
           if (!datasetName) return;
 
-          const controller = new AbortController();
-          _datasetCheckController = controller;
-          set({ isCheckingDataset: true });
+          runDatasetCheck(datasetName, datasetSplit || "train");
+        },
+        ensureDatasetChecked: () => {
+          const state = get();
+          if (state.isCheckingDataset) return;
+          if (state.isDatasetMultimodal !== null) return;
 
-          checkDatasetFormat({
-            datasetName,
-            hfToken: state.hfToken.trim() || null,
-            subset: state.datasetSubset,
-            split: datasetSplit || "train",
-          })
-            .then((res) => {
-              if (controller.signal.aborted) return;
-              set({
-                isDatasetMultimodal: !!res.is_multimodal,
-                isCheckingDataset: false,
-              });
-            })
-            .catch(() => {
-              if (controller.signal.aborted) return;
-              set({ isDatasetMultimodal: null, isCheckingDataset: false });
-            });
+          const datasetName =
+            state.datasetSource === "huggingface"
+              ? state.dataset
+              : state.uploadedFile;
+          if (!datasetName) return;
+
+          const split = state.datasetSplit || "train";
+          runDatasetCheck(datasetName, split);
         },
         setDatasetManualMapping: (datasetManualMapping) =>
           set({ datasetManualMapping }),
@@ -274,8 +318,10 @@ export const useTrainingConfigStore = create<TrainingConfigStore>()(
         setSaveSteps: (saveSteps) => set({ saveSteps }),
         setEvalSteps: (evalSteps) => set({ evalSteps }),
         setPacking: (packing) => set({ packing }),
-        setTrainOnCompletions: (trainOnCompletions) =>
-          set({ trainOnCompletions }),
+        setTrainOnCompletions: (trainOnCompletions) => {
+          _trainOnCompletionsManuallySet = true;
+          set({ trainOnCompletions });
+        },
         setGradientCheckpointing: (gradientCheckpointing) =>
           set({ gradientCheckpointing }),
         setRandomSeed: (randomSeed) => set({ randomSeed }),
diff --git a/studio/frontend/src/features/training/types/config.ts b/studio/frontend/src/features/training/types/config.ts
index 67f1d00edc..53beabd383 100644
--- a/studio/frontend/src/features/training/types/config.ts
+++ b/studio/frontend/src/features/training/types/config.ts
@@ -72,6 +72,7 @@ export interface TrainingConfigActions {
   setModelType: (type: ModelType) => void;
   setSelectedModel: (model: string | null) => void;
   ensureModelDefaultsLoaded: () => void;
+  ensureDatasetChecked: () => void;
   setTrainingMethod: (method: TrainingMethod) => void;
   setHfToken: (token: string) => void;
   setDatasetSource: (source: DatasetSource) => void;