chore(skills): shorten eval smoke prompts

Signed-off-by: Chen Cui <chcui@nvidia.com>

Author: cuichenx

skills/nemo-mbridge-mlm-bridge-training/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "mlm-bridge-training-positive-recipe-smoke",
-    "question": "I need to run a Megatron Bridge training smoke test with mock data on multiple GPUs and compare it against Megatron-LM behavior. Which training entry points and recipe patterns should I use?",
+    "question": "I need to run a Megatron Bridge training smoke test with mock data on multiple GPUs and compare it against Megatron-LM behavior. Which training entry points and recipe patterns should I use? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-mlm-bridge-training",
     "expected_script": null,
     "ground_truth": "The answer should route to Megatron Bridge/Megatron-LM training guidance, describe using the repo's training recipes or examples with mock data for smoke/correlation testing, and point the user toward multi-GPU launch patterns. It should distinguish Bridge recipes from direct Megatron-LM submodule changes and warn that the pinned Megatron-Core submodule should not be edited directly.",

skills/nemo-mbridge-multi-node-slurm/evals/evals.json

@@ -1,16 +1,17 @@
 [
   {
     "id": "multi-node-slurm-positive-sbatch-smoke",
-    "question": "Convert my single-node Megatron Bridge launch into a two-node Slurm sbatch job. I need guidance on srun-native launch, container mounts, NCCL variables, and debugging timeouts.",
+    "question": "For a Megatron Bridge recipe that goes through initialize.py, convert my single-node launch to a two-node Slurm sbatch plan. Include the preferred srun-native shape, shared cache/mount requirements, and the exact first log checks for NCCL timeout debugging. Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-multi-node-slurm",
     "expected_script": null,
-    "ground_truth": "The answer should use the multi-node Slurm skill, recommend the srun-native approach as the preferred pattern, explain cluster/container setup considerations, and include practical checks for NCCL timeouts and multi-node debugging. It should not default to a local single-node torchrun-only recipe.",
+    "ground_truth": "The answer should use the multi-node Slurm skill and recommend the Bridge srun-native pattern: Slurm launches 8 tasks per node, not torch.distributed.run spawning inside one Slurm task. It should state that Bridge derives RANK, WORLD_SIZE, LOCAL_RANK, MASTER_ADDR, and MASTER_PORT from SLURM env vars, require shared filesystem paths for repo/data/logs/HF_HOME/UV_CACHE_DIR/NEMO_HOME plus container mounts, and give the first timeout-debugging checks: grep for real errors while filtering noise, inspect the first failed rank/node, and check NCCL ncclUniqueId/timeout or rank-0 crash lines.",
     "expected_behavior": [
       "Read the nemo-mbridge-multi-node-slurm skill before answering.",
       "Identify the task as multi-node Slurm launch conversion.",
-      "Recommend the srun-native approach unless the user explicitly needs the legacy torch.distributed path.",
-      "Mention container, mount, and environment setup considerations.",
-      "Include debugging advice for NCCL or node-communication failures."
+      "Recommend the srun-native Bridge approach with Slurm spawning 8 tasks per node.",
+      "Mention that Bridge derives distributed rank and rendezvous variables from SLURM env vars.",
+      "Require shared cache/storage paths and container mounts for multi-node jobs.",
+      "List the first NCCL timeout debugging checks: filtered error grep, first failed rank/node, and ncclUniqueId or rank-0 crash lines."
     ]
   }
 ]

skills/nemo-mbridge-perf-activation-recompute/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "activation-recompute-positive-memory-smoke",
-    "question": "My Megatron Bridge model is close to OOM. Should I enable selective activation recompute or full activation recompute, and what tradeoff should I expect?",
+    "question": "My Megatron Bridge model is close to OOM. Should I enable selective activation recompute or full activation recompute, and what tradeoff should I expect? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-activation-recompute",
     "expected_script": null,
     "ground_truth": "The answer should explain that activation recompute reduces GPU memory by recomputing activations during backward, at the cost of extra compute. It should recommend selective recompute first when possible, reserve full recompute for tighter memory pressure, and note interactions with other features such as CUDA graphs and parallelism choices.",

skills/nemo-mbridge-perf-cpu-offloading/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "cpu-offloading-positive-optimizer-smoke",
-    "question": "For a Qwen3-30B-A3B Megatron Bridge run that needs pipeline parallelism and still has GPU memory pressure, should I use activation CPU offloading or optimizer CPU offloading? Include the exact optimizer config knobs and the main activation-offload constraints.",
+    "question": "For a Qwen3-30B-A3B Megatron Bridge run that needs pipeline parallelism and still has GPU memory pressure, should I use activation CPU offloading or optimizer CPU offloading? Include the exact optimizer config knobs and the main activation-offload constraints. Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-cpu-offloading",
     "expected_script": null,
     "ground_truth": "The answer should use the CPU offloading skill, choose optimizer CPU offloading for large MoE models that need pipeline parallelism, and explain that layer-level activation CPU offloading requires pipeline_model_parallel_size=1. It should include optimizer.optimizer_cpu_offload=True, optimizer.optimizer_offload_fraction, and optionally optimizer.overlap_cpu_optimizer_d2h_h2d=True. It should mention activation offloading constraints: PP=1, no activation recompute, no CUDA graphs, and cpu_offloading_num_layers in range.",

skills/nemo-mbridge-perf-cuda-graphs/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "cuda-graphs-te-scoped-moe-smoke",
-    "question": "I am training a Megatron Bridge Qwen3 MoE model and want to reduce CPU launch overhead with CUDA graphs. Which cuda_graph_impl and cuda_graph_scope should I start with, and what prerequisites should I set?",
+    "question": "I am training a Megatron Bridge Qwen3 MoE model and want to reduce CPU launch overhead with CUDA graphs. Which cuda_graph_impl and cuda_graph_scope should I start with, and what prerequisites should I set? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-cuda-graphs",
     "expected_script": null,
     "ground_truth": "For most Megatron Bridge training workloads, start with Transformer Engine scoped graphs rather than local full-iteration capture. For a dropless MoE model, use cuda_graph_impl=\"transformer_engine\" with cuda_graph_scope including attn, moe_router, and moe_preprocess. Set cuda_graph_warmup_steps, enable model.use_te_rng_tracker and rng.te_rng_tracker, keep sequence length and micro-batch size static, and compare steady-state replay iterations after warmup and capture. Do not combine moe and moe_router scopes.",

skills/nemo-mbridge-perf-expert-parallel-overlap/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "expert-parallel-overlap-positive-flex-smoke",
-    "question": "For a Megatron Bridge MoE model using expert parallelism, how do I enable expert-parallel communication overlap with DeepEP or HybridEP, and what should I verify?",
+    "question": "For a Megatron Bridge MoE model using expert parallelism, how do I enable expert-parallel communication overlap with DeepEP or HybridEP, and what should I verify? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-expert-parallel-overlap",
     "expected_script": null,
     "ground_truth": "The answer should route to expert-parallel communication overlap guidance, discuss overlap_moe_expert_parallel_comm and related delayed expert wgrad scheduling, and call out dispatcher/backend choices such as DeepEP or HybridEP. It should include compatibility checks before enabling overlap and advise comparing short-run measurements to the baseline.",

skills/nemo-mbridge-perf-hierarchical-context-parallel/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "hierarchical-context-parallel-positive-long-context-smoke",
-    "question": "I want to train a long-context Megatron Bridge model using hierarchical context parallelism. Which config knobs and verification checks should I look at?",
+    "question": "I want to train a long-context Megatron Bridge model using hierarchical context parallelism. Which config knobs and verification checks should I look at? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-hierarchical-context-parallel",
     "expected_script": null,
     "ground_truth": "The answer should focus on hierarchical context parallelism, describe enabling the relevant context-parallel hierarchy knobs at a high level, and direct the user to verify group sizing, sequence-length divisibility, communication topology, and runtime logs. It should distinguish hierarchical CP from ordinary tensor or data parallel sizing.",

skills/nemo-mbridge-perf-megatron-fsdp/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "megatron-fsdp-positive-enable-smoke",
-    "question": "How should I enable Megatron FSDP in Megatron Bridge, and what pitfalls should I check before running a large model?",
+    "question": "How should I enable Megatron FSDP in Megatron Bridge, and what pitfalls should I check before running a large model? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-megatron-fsdp",
     "expected_script": null,
     "ground_truth": "The answer should use Megatron FSDP guidance, summarize the relevant enablement/configuration flow, and call out pitfalls such as compatibility with parallelism choices, memory expectations, and verification from logs or config. It should not substitute generic PyTorch FSDP advice without mapping back to Megatron Bridge.",

skills/nemo-mbridge-perf-memory-tuning/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "memory-tuning-positive-oom-smoke",
-    "question": "My Megatron Bridge training run is OOM on H100. How should I systematically reduce peak GPU memory before changing the model?",
+    "question": "My Megatron Bridge training run is OOM on H100. How should I systematically reduce peak GPU memory before changing the model? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-memory-tuning",
     "expected_script": null,
     "ground_truth": "The answer should use the memory tuning skill and propose a systematic memory-reduction path: verify the eager baseline and memory profile, consider expandable segments, adjust parallelism or micro-batch sizing, use activation recompute, consider CPU offloading only when appropriate, and check known incompatibilities. It should avoid jumping directly to a single feature without diagnosis.",

skills/nemo-mbridge-perf-moe-comm-overlap/evals/evals.json

@@ -1,7 +1,7 @@
 [
   {
     "id": "moe-comm-overlap-positive-dispatch-combine-smoke",
-    "question": "For my Megatron Bridge MoE run, how do I overlap expert-parallel dispatch and combine communication, and how does this interact with CUDA graphs or recompute?",
+    "question": "For my Megatron Bridge MoE run, how do I overlap expert-parallel dispatch and combine communication, and how does this interact with CUDA graphs or recompute? Keep the answer to at most five bullets.",
     "expected_skill": "nemo-mbridge-perf-moe-comm-overlap",
     "expected_script": null,
     "ground_truth": "The answer should use MoE communication-overlap guidance, describe dispatch/combine overlap and expert wgrad scheduling at a high level, and explicitly discuss interactions with recompute and CUDA graphs. It should recommend measuring a short run against the non-overlap baseline.",