feat(grpo): optional early stop at a validation metric threshold

examples/configs/grpo_math_1B.yaml

@@ -15,6 +15,10 @@ grpo:
   advantage_clip_low: null
   advantage_clip_high: null
   max_val_samples: 256
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
   use_dynamic_sampling: false

examples/nemo_gym/grpo_nanov3.yaml

@@ -16,6 +16,10 @@ grpo:
   advantage_clip_low: null
   advantage_clip_high: null
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
   async_grpo:

examples/nemo_gym/grpo_workplace_assistant_nemotron_nano_v2_9b.yaml

@@ -14,6 +14,10 @@ grpo:
   advantage_clip_low: null
   advantage_clip_high: null
   max_val_samples: null   # inferred from size of val dataset. for multi evals, repeat val ds via `num_repeats` in `ng_prepare_data`.
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: null
   seed: 42
   use_dynamic_sampling: false

examples/nemo_gym/nemotron-3-super/stage1_rlvr.yaml

@@ -27,6 +27,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-super/stage2_swe1.yaml

@@ -27,6 +27,10 @@ grpo:
   val_at_end: false
   overlong_filtering: true
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-super/stage2_swe2.yaml

@@ -27,6 +27,10 @@ grpo:
   val_at_end: false
   overlong_filtering: true
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-super/stage3_rlhf.yaml

@@ -27,6 +27,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/ifbench_teacher.yaml

@@ -61,6 +61,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/mopd.yaml

@@ -76,6 +76,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/reasoning_teacher.yaml

@@ -65,6 +65,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/rlhf_teacher.yaml

@@ -62,6 +62,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/student_rlvr1.yaml

@@ -58,6 +58,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/student_rlvr2.yaml

@@ -59,6 +59,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42
 

examples/nemo_gym/nemotron-3-ultra/swe_teacher.yaml

@@ -81,6 +81,10 @@ grpo:
   val_at_end: false
   overlong_filtering: false
   max_val_samples: null
+  # Early stop once this metric (e.g. accuracy or pass_k) reaches the threshold; null disables.
+  stop_at_validation_metric: null
+  # Required when stop_at_validation_metric is set.
+  stop_at_validation_threshold: null
   val_batch_size: 256
   seed: 42