diff --git a/examples/mlperf/model_train.py b/examples/mlperf/model_train.py
index ed4b33a77a..2cccfff804 100644
--- a/examples/mlperf/model_train.py
+++ b/examples/mlperf/model_train.py
@@ -398,11 +398,11 @@ def train_bert():
   seed = config["seed"] = getenv("SEED", 12345)
 
   # ** hyperparameters **
-  BS                 = config["GLOBAL_BATCH_SIZE"]      = getenv("BS", 4 * len(GPUS)) # FP32 4090: 6 GPUS -> BS24
-  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 4 * len(GPUS))
-  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.000004166 * BS)
+  BS                 = config["GLOBAL_BATCH_SIZE"]      = getenv("BS", 16 * len(GPUS) if dtypes.default_float in (dtypes.float16, dtypes.bfloat16) else 8 * len(GPUS))
+  EVAL_BS            = config["EVAL_BS"]                = getenv("EVAL_BS", 1 * len(GPUS))
+  max_lr             = config["OPT_BASE_LEARNING_RATE"] = getenv("OPT_BASE_LEARNING_RATE", 0.00035 * math.sqrt(BS/256))
 
-  train_steps        = config["TRAIN_STEPS"]            = getenv("TRAIN_STEPS", 3840000 // BS)
+  train_steps        = config["TRAIN_STEPS"]            = getenv("TRAIN_STEPS", 4800000 // BS)
   warmup_steps       = config["NUM_WARMUP_STEPS"]       = getenv("NUM_WARMUP_STEPS", 1)
   max_eval_steps     = config["MAX_EVAL_STEPS"]         = getenv("MAX_EVAL_STEPS", (10000 + EVAL_BS - 1) // EVAL_BS) # EVAL_BS * MAX_EVAL_STEPS >= 10000
   eval_step_freq     = config["EVAL_STEP_FREQ"]         = getenv("EVAL_STEP_FREQ", int((math.floor(0.05 * (230.23 * BS + 3000000) / 25000) * 25000) / BS)) # Round down