feat(grpo): validation-only sampling params and grouped pass@k validation

examples/configs/distillation_math.yaml

@@ -201,6 +201,9 @@ policy: &POLICY_BASE
         temperature: 1.0
         top_p: 1.0
         top_k: null
+        val_temperature: ${.temperature}
+        val_top_p: ${.top_p}
+        val_top_k: ${.top_k}
         stop_token_ids: null
         stop_strings: null
         vllm_cfg:

examples/configs/evals/eval.yaml

@@ -12,6 +12,9 @@ generation:
   temperature: 0.0
   top_p: 1.0
   top_k: -1 # -1 means disable
+  val_temperature: ${.temperature}
+  val_top_p: ${.top_p}
+  val_top_k: ${.top_k}
   num_prompts_per_step: -1 # -1 means pass all prompts at once
   model_name: "Qwen/Qwen2.5-Math-1.5B-Instruct"
   stop_token_ids: null

examples/configs/evals/mmau.yaml

@@ -11,6 +11,9 @@ generation:
   temperature: 0.0
   top_p: 1.0
   top_k: -1
+  val_temperature: ${.temperature}
+  val_top_p: ${.top_p}
+  val_top_k: ${.top_k}
   num_prompts_per_step: -1
   model_name: "Qwen/Qwen2.5-Omni-3B"
   stop_token_ids: null

examples/configs/grpo_math_1B.yaml

@@ -15,6 +15,9 @@ grpo:
   advantage_clip_low: null
   advantage_clip_high: null
   max_val_samples: 256
+  # Validation rollouts per prompt; k > 1 also reports the pass_k metric.
+  # max_val_samples counts PROMPTS: total validation rollouts = max_val_samples * k.
+  val_num_generations_per_prompt: 1
   # 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.
@@ -344,6 +347,10 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    # Validation-only sampling; defaults follow the train values above.
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # null = topology default (IPC colocated, NCCL non-colocated).

examples/configs/ppo_math_1B.yaml

@@ -234,6 +234,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     mcore_generation_config:

examples/configs/recipes/llm/mopd-qwen3-1.7b-3n8g-megatron-pack.yaml

@@ -2,7 +2,6 @@ defaults: ../../grpo_math_1B.yaml
 grpo:
   num_prompts_per_step: 8
   num_generations_per_prompt: 4
-  num_val_generations_per_prompt: 1
   max_num_steps: 5
   val_period: 1000
   overlong_filtering: true

examples/nemo_gym/grpo_nanov3.yaml

@@ -2,7 +2,7 @@
 grpo:
   num_prompts_per_step: 128
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 4
+  val_num_generations_per_prompt: 4
   max_rollout_turns: 1 # for multi-turn rollouts. Math Environments just have 1 turn (answering the question)
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -210,6 +210,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     mcore_generation_config:

examples/nemo_gym/grpo_qwen3_30ba3b_thinking_swe1.yaml

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 64
   num_generations_per_prompt: 8
-  num_val_generations_per_prompt: 1
+  val_num_generations_per_prompt: 1
   max_num_epochs: 100
   advantage_clip_low: -100
   advantage_clip_high: 100

examples/nemo_gym/grpo_qwen3_30ba3b_thinking_swe2.yaml

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 8
   num_generations_per_prompt: 8
-  num_val_generations_per_prompt: 1
+  val_num_generations_per_prompt: 1
   max_num_epochs: 100
   advantage_clip_low: -100
   advantage_clip_high: 100

examples/nemo_gym/grpo_workplace_assistant_nemotron_nano_v2_9b.yaml

@@ -14,6 +14,7 @@ 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`.
+  val_num_generations_per_prompt: 1
   # 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.
@@ -226,6 +227,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg:

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

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 256
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -226,6 +226,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg:

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

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 64
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 1
+  val_num_generations_per_prompt: 1
   max_rollout_turns: 1
   max_num_epochs: 100
   max_num_steps: 1000000
@@ -226,6 +226,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg:

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

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 16
   num_generations_per_prompt: 32
-  num_val_generations_per_prompt: 1
+  val_num_generations_per_prompt: 1
   max_rollout_turns: 1
   max_num_epochs: 100
   max_num_steps: 1000000
@@ -219,6 +219,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg:

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

@@ -13,7 +13,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 128
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -226,6 +226,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg:

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

@@ -47,7 +47,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 128
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -305,6 +305,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -62,7 +62,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 1024
   num_generations_per_prompt: 1
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -326,6 +326,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -51,7 +51,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 128
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 10
   max_num_steps: 1000000
@@ -308,6 +308,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -48,7 +48,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 128
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -306,6 +306,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -44,7 +44,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 512
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -302,6 +302,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -45,7 +45,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 512
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 1
   max_num_steps: 1000000
@@ -303,6 +303,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     # TP=8 EP=8: EP=TP so vllm_dp_size=1, async_engine=true works with NeMo Gym.

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

@@ -67,7 +67,7 @@ checkpointing:
 grpo:
   num_prompts_per_step: 32
   num_generations_per_prompt: 16
-  num_val_generations_per_prompt: 2
+  val_num_generations_per_prompt: 2
   max_rollout_turns: 1
   max_num_epochs: 4
   max_num_steps: 1000000
@@ -325,6 +325,9 @@ policy:
     temperature: 1.0
     top_p: 1.0
     top_k: null
+    val_temperature: ${.temperature}
+    val_top_p: ${.top_p}
+    val_top_k: ${.top_k}
     stop_token_ids: null
     stop_strings: null
     vllm_cfg: