[Quality] GLM-5.2 quality benchmarks (#36) merged with vng-benchmark

.github/workflows/benchmark-tmpl.yml

@@ -15,6 +15,9 @@
       # recipe ignores it, and the SPEED-Bench path falls back to HF_TOKEN.
       SPEEDBENCH_HF_TOKEN:
         required: false
+      # Bearer token for the GreenNode MaaS endpoint (quality evals, remote bench).
+      GREENNODE_API_KEY:
+        required: false
     inputs:
       config:
         description: "Benchmark configuration as JSON"
@@ -173,28 +176,28 @@
   RECIPE_FINGERPRINT: ${{ fromJSON(inputs.config).recipe-fingerprint || '' }}
   MODEL: ${{ fromJSON(inputs.config).model }}
   MODEL_PREFIX: ${{ fromJSON(inputs.config).model-prefix }}
-  ISL: ${{ inputs.scenario-type == 'agentic-coding' && '0' || fromJSON(inputs.config).isl }}
-  OSL: ${{ inputs.scenario-type == 'agentic-coding' && '0' || fromJSON(inputs.config).osl }}
-  MAX_MODEL_LEN: ${{ inputs.scenario-type == 'agentic-coding' && '0' || fromJSON(inputs.config).max-model-len }}
+  ISL: ${{ (inputs.scenario-type == 'agentic-coding' || startsWith(inputs.scenario-type, 'quality-')) && '0' || fromJSON(inputs.config).isl }}
+  OSL: ${{ (inputs.scenario-type == 'agentic-coding' || startsWith(inputs.scenario-type, 'quality-')) && '0' || fromJSON(inputs.config).osl }}
+  MAX_MODEL_LEN: ${{ (inputs.scenario-type == 'agentic-coding' || startsWith(inputs.scenario-type, 'quality-')) && '0' || fromJSON(inputs.config).max-model-len }}
   IMAGE: ${{ fromJSON(inputs.config).image }}
   FRAMEWORK: ${{ fromJSON(inputs.config).framework }}
   PRECISION: ${{ fromJSON(inputs.config).precision }}
-  TP: ${{ fromJSON(inputs.config).tp }}
-  PP_SIZE: ${{ fromJSON(inputs.config).pp }}
-  DCP_SIZE: ${{ fromJSON(inputs.config).dcp-size }}
-  PCP_SIZE: ${{ fromJSON(inputs.config).pcp-size }}
-  EP_SIZE: ${{ fromJSON(inputs.config).ep }}
+  TP: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).tp }}
+  PP_SIZE: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).pp }}
+  DCP_SIZE: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).dcp-size }}
+  PCP_SIZE: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).pcp-size }}
+  EP_SIZE: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).ep }}
   DP_ATTENTION: ${{ inputs.dp-attn }}
-  CONC: ${{ fromJSON(inputs.config).conc }}
-  SPEC_DECODING: ${{ fromJSON(inputs.config).spec-decoding }}
-  DISAGG: ${{ inputs.scenario-type == 'agentic-coding' && 'false' || fromJSON(inputs.config).disagg }}
+  CONC: ${{ startsWith(inputs.scenario-type, 'quality-') && '1' || fromJSON(inputs.config).conc }}
+  SPEC_DECODING: ${{ startsWith(inputs.scenario-type, 'quality-') && 'none' || fromJSON(inputs.config).spec-decoding }}
+  DISAGG: ${{ (inputs.scenario-type == 'agentic-coding' || startsWith(inputs.scenario-type, 'quality-')) && 'false' || fromJSON(inputs.config).disagg }}
   RUN_EVAL: ${{ inputs.run-eval }}
   EVAL_ONLY: ${{ inputs.eval-only }}
   EVAL_FRAMEWORK: ${{ inputs.eval-framework }}
   EVAL_SUITE: ${{ inputs.eval-suite }}
   # Agentic-coding env. Fixed-seq-len jobs leave these empty.
   SCENARIO_TYPE: ${{ inputs.scenario-type }}
-  SCENARIO_SUBDIR: ${{ inputs.scenario-type == 'agentic-coding' && 'agentic/' || 'fixed_seq_len/' }}
+  SCENARIO_SUBDIR: ${{ (inputs.scenario-type == 'agentic-coding' && 'agentic/') || (startsWith(inputs.scenario-type, 'quality-') && 'quality/') || 'fixed_seq_len/' }}
   IS_AGENTIC: ${{ inputs.scenario-type == 'agentic-coding' && '1' || '0' }}
   KV_OFFLOADING: ${{ inputs.scenario-type == 'agentic-coding' && format('{0}', fromJSON(inputs.config).kv-offloading) || '' }}
   KV_OFFLOAD_BACKEND: ${{ inputs.scenario-type == 'agentic-coding' && format('{0}', fromJSON(inputs.config).kv-offload-backend.name) || '' }}
@@ -227,11 +230,11 @@
   # scenario in one run). scenario-map is keyed by the row's exp-name; the flat
   # serving.scenario is the single-scenario fallback, empty for scenario lists
   # so a missing map entry fails loudly in serve_generic.sh (check_env_vars).
   SERVING_SCENARIO: ${{ inputs.serving != '' && (fromJSON(inputs.serving).scenario-map[fromJSON(inputs.config).exp-name].scenario || fromJSON(inputs.serving).scenario) || '' }}
   SERVING_LOADER: ${{ inputs.serving != '' && (fromJSON(inputs.serving).scenario-map[fromJSON(inputs.config).exp-name].loader || fromJSON(inputs.serving).loader) || '' }}
   # '1' for sweep scenarios (adaptive_search), '0' otherwise — never '': check_env_vars
   # rejects empty values, so '' would abort every non-sweep serving run in serve_generic.sh.
   SERVING_SEARCH: ${{ inputs.serving != '' && ((fromJSON(inputs.serving).scenario-map[fromJSON(inputs.config).exp-name].search || fromJSON(inputs.serving).search) && '1' || '0') }}
   SERVING_YAML: ${{ inputs.serving != '' && fromJSON(inputs.serving).yaml || '' }}
   # Router sidecar (any router, spec 2026-10-14); only runners/launch_hardware-hcm.sh reads these.
   ROUTER_IMAGE: ${{ inputs.serving != '' && fromJSON(inputs.serving).router-image || '' }}
@@ -247,6 +250,14 @@
   # Bearer token for authenticated remote targets. Empty for every other
   # recipe, which leaves build_replay_cmd's --api-key branch inert.
   REMOTE_API_KEY: ${{ secrets.GREENNODE_API_KEY }}
+  # Quality-eval env, from the quality row (scenario-type quality-*); empty for every other recipe.
+  QUALITY_BENCHMARK_NAME: ${{ startsWith(inputs.scenario-type, 'quality-') && fromJSON(inputs.config).benchmark-name || '' }}
+  QUALITY_ENDPOINT: ${{ startsWith(inputs.scenario-type, 'quality-') && fromJSON(inputs.config).quality-endpoint || '' }}
+  QUALITY_API_KEY: ${{ secrets.GREENNODE_API_KEY }}
+  QUALITY_MODEL_NAME: ${{ startsWith(inputs.scenario-type, 'quality-') && fromJSON(inputs.config).quality-model-name || '' }}
+  SMOKE: ${{ startsWith(inputs.scenario-type, 'quality-') && fromJSON(inputs.config).smoke && 'true' || 'false' }}
+  NUM_CONCURRENT: ${{ startsWith(inputs.scenario-type, 'quality-') && format('{0}', fromJSON(inputs.config).num-concurrent) || '' }}
+  MAX_GEN_TOKENS: ${{ startsWith(inputs.scenario-type, 'quality-') && format('{0}', fromJSON(inputs.config).max-gen-tokens) || '' }}
   AIPERF_FAILED_REQUEST_THRESHOLD: '0.10'
   RESULT_DIR: /workspace/results
   PYTHONDONTWRITEBYTECODE: '1'
@@ -261,7 +272,7 @@
 
 jobs:
   benchmark:
     runs-on: >-
       ${{ fromJSON(
         vars.PRIORITY_SCHEDULER_ENABLED == 'true' &&
         (
@@ -306,6 +317,7 @@
     timeout-minutes: 500
     name: >-
       ${{ inputs.klaud-run && 'klaud | ' || '' }}p${{ inputs.priority }} | ${{ fromJSON(inputs.config).model-prefix }} ${{ fromJSON(inputs.config).precision }} ${{ inputs.runner }} ${{ fromJSON(inputs.config).framework == 'sglang' && 'sgl' || fromJSON(inputs.config).framework == 'dynamo-sglang' && 'dyn-sgl' || fromJSON(inputs.config).framework == 'sglang-disagg' && 'sgl-disagg' || fromJSON(inputs.config).framework }}
+      ${{ startsWith(inputs.scenario-type, 'quality-') && fromJSON(inputs.config).benchmark-name || '' }}
       TP${{ fromJSON(inputs.config).tp }}${{ format('{0}', fromJSON(inputs.config).pp) != '' && format('{0}', fromJSON(inputs.config).pp) != '1' && format('/PP{0}', fromJSON(inputs.config).pp) || '' }}${{ format('{0}', fromJSON(inputs.config).dcp-size) != '' && format('{0}', fromJSON(inputs.config).dcp-size) != '1' && format('/DCP{0}', fromJSON(inputs.config).dcp-size) || '' }}${{ format('{0}', fromJSON(inputs.config).pcp-size) != '' && format('{0}', fromJSON(inputs.config).pcp-size) != '1' && format('/PCP{0}', fromJSON(inputs.config).pcp-size) || '' }}${{ format('{0}', fromJSON(inputs.config).ep) != '' && format('{0}', fromJSON(inputs.config).ep) != '1' && format('/EP{0}', fromJSON(inputs.config).ep) || '' }}${{ inputs.dp-attn && '/DPA' || '' }}
       ${{ fromJSON(inputs.config).spec-decoding != 'none' && fromJSON(inputs.config).spec-decoding || '' }}
       ${{ inputs.scenario-type == 'agentic-coding' && format('{0}', fromJSON(inputs.config).kv-offloading) != '' && fromJSON(inputs.config).kv-offloading != 'none' && format('{0} KV offload', fromJSON(inputs.config).kv-offloading) || '' }}
@@ -440,6 +452,11 @@
             hardware-hcm-*) LAUNCH_SCRIPT="runners/launch_hardware-hcm.sh" ;;
             *) LAUNCH_SCRIPT="runners/launch_${RUNNER_NAME%%_*}.sh" ;;
           esac
+          # Quality evals drive a remote endpoint from any runner labelled
+          # cluster:quality-eval, so they skip the per-runner launchers.
+          if [[ "$SCENARIO_SUBDIR" == "quality/" ]]; then
+            LAUNCH_SCRIPT="runners/launch_quality-eval.sh"
+          fi
           bash "$LAUNCH_SCRIPT"
 
           if [ "${EVAL_ONLY}" = "true" ]; then
@@ -590,12 +607,24 @@
             agent_preds.json
             predictions.jsonl
             swebench_report_*.json
+            eval_results*.json
             *.traj*
+            *.csv
+            lcb_results*.json
+            *.jsonl
+            bfcl_inference_audit.json
+            bfcl_result_*.json
+            scicode_debug.tar.gz
           if-no-files-found: ${{ inputs.eval-only && 'error' || 'ignore' }}
 
       - name: Verify eval scores
         if: ${{ (success() || failure()) && inputs.eval-only }}
-        run: python3 utils/evals/validate_scores.py
+        run: |
+          SMOKE_FLAG=""
+          if [ "$SMOKE" = "true" ]; then
+            SMOKE_FLAG="--smoke"
+          fi
+          python3 utils/evals/validate_scores.py $SMOKE_FLAG
 
       - name: Cleanup eval outputs (post-upload)
         if: ${{ always() && (env.RUN_EVAL == 'true' || inputs.eval-only) }}
@@ -608,6 +637,8 @@
           rm -f sample*.jsonl || true
           rm -f -- ./*_artifacts.tar.gz || true
           rm -f agent_preds.json predictions.jsonl swebench_report_*.json *.traj* || true
+          rm -f eval_results*.json *.csv lcb_results*.json || true
+          rm -f bfcl_inference_audit.json bfcl_result_*.json scicode_debug.tar.gz || true
 
       - name: Repair root-owned artifacts (post-run)
         if: ${{ always() && inputs.runner == 'cluster:mi355x-amds' }}

.github/workflows/e2e-tests.yml

@@ -231,6 +231,7 @@ jobs:
             agentic-eval-config: ${{ steps.get-jobs.outputs.agentic-eval-config }}
             multi-node-agentic-config: ${{ steps.get-jobs.outputs.multi-node-agentic-config }}
             multi-node-agentic-eval-config: ${{ steps.get-jobs.outputs.multi-node-agentic-eval-config }}
+            quality-eval-config: ${{ steps.get-jobs.outputs.quality-eval-config }}
             serving: ${{ steps.get-jobs.outputs.serving }}
         steps:
             - name: Checkout code (ref)
@@ -349,22 +350,50 @@ jobs:
                   AGENTIC_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' not in x and x.get('run-eval', False)]))" | score_matrix agentic-eval)
                   MULTI_AGENTIC=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' in x and not x.get('eval-only', False)]))" | score_matrix multi-agentic)
                   MULTI_AGENTIC_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type') == 'agentic-coding' and 'prefill' in x and x.get('run-eval', False)]))" | score_matrix multi-agentic-eval)
-                  SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix single)
-                  MULTI=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and not x.get('eval-only', False)]))" | score_matrix multi)
-                  EVALS=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix eval)
+                  QUALITY=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if x.get('scenario-type', '').startswith('quality-')]))" | score_matrix quality-eval)
+                  SINGLE=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('scenario-type', '').startswith('quality-') and not x.get('eval-only', False)]))" | score_matrix single)
+                  MULTI=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and not x.get('scenario-type', '').startswith('quality-') and not x.get('eval-only', False)]))" | score_matrix multi)
+                  EVALS=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' not in x and x.get('scenario-type') != 'agentic-coding' and not x.get('scenario-type', '').startswith('quality-') and x.get('run-eval', False)]))" | score_matrix eval)
                   MULTI_EVAL=$(echo "$CONFIG_JSON" | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps([x for x in d if 'prefill' in x and x.get('scenario-type') != 'agentic-coding' and x.get('run-eval', False)]))" | score_matrix multi-eval)
                   {
                     echo "agentic-config=$AGENTIC"
                     echo "agentic-eval-config=$AGENTIC_EVAL"
                     echo "multi-node-agentic-config=$MULTI_AGENTIC"
                     echo "multi-node-agentic-eval-config=$MULTI_AGENTIC_EVAL"
+                    echo "quality-eval-config=$QUALITY"
                     echo "single-node-config=$SINGLE"
                     echo "multi-node-config=$MULTI"
                     echo "eval-config=$EVALS"
                     echo "multi-node-eval-config=$MULTI_EVAL"
                     echo "serving=${SERVING_JSON:-}"
                   } >> "$GITHUB_OUTPUT"
 
+    test-sweep-quality-evals:
+        needs: get-jobs
+        if: ${{ needs.get-jobs.outputs.quality-eval-config != '' && needs.get-jobs.outputs.quality-eval-config != '[]' }}
+        uses: $/.github/workflows/benchmark-tmpl.yml
+        name: quality eval /
+        strategy:
+            fail-fast: ${{ inputs.fail-fast }}
+            matrix:
+                config: ${{ fromJson(needs.get-jobs.outputs.quality-eval-config) }}
+        secrets:
+            REPO_PAT: ${{ secrets.REPO_PAT }}
+            INFERENCEX_OFFICIAL_RO_HF_TOKEN: ${{ secrets.INFERENCEX_OFFICIAL_RO_HF_TOKEN }}
+            GREENNODE_API_KEY: ${{ secrets.GREENNODE_API_KEY }}
+        with:
+            config: ${{ toJSON(matrix.config) }}
+            klaud-run: ${{ inputs.klaud-run }}
+            runner: ${{ matrix.config.runner }}
+            priority: ${{ matrix.config.priority }}
+            queue-token: ${{ matrix.config['queue-token'] }}
+            skip-queue-pr: ${{ matrix.config['skip-queue-pr'] || '' }}
+            run-eval: true
+            eval-only: true
+            eval-limit: ${{ matrix.config.eval-limit || inputs.eval-limit }}
+            scenario-type: ${{ matrix.config.scenario-type }}
+            ref: ${{ inputs.ref }}
+
     test-sweep-multi-node:
         needs: get-jobs
         if: ${{ needs.get-jobs.outputs.multi-node-config != '' && needs.get-jobs.outputs.multi-node-config != '[]' }}
@@ -617,8 +646,8 @@ jobs:
 
     collect-evals:
         name: collect-evals
-        needs: [test-sweep-evals, test-sweep-multi-node-evals, test-sweep-agentic-evals, test-sweep-multi-node-agentic-evals]
-        if: ${{ always() && (needs.test-sweep-evals.result != 'skipped' || needs.test-sweep-multi-node-evals.result != 'skipped' || needs.test-sweep-agentic-evals.result != 'skipped' || needs.test-sweep-multi-node-agentic-evals.result != 'skipped') }}
+        needs: [test-sweep-evals, test-sweep-multi-node-evals, test-sweep-agentic-evals, test-sweep-multi-node-agentic-evals, test-sweep-quality-evals]
+        if: ${{ always() && (needs.test-sweep-evals.result != 'skipped' || needs.test-sweep-multi-node-evals.result != 'skipped' || needs.test-sweep-agentic-evals.result != 'skipped' || needs.test-sweep-multi-node-agentic-evals.result != 'skipped' || needs.test-sweep-quality-evals.result != 'skipped') }}
         uses: $/.github/workflows/collect-evals.yml
 
     calc-success-rate:

.github/workflows/run-sweep.yml

@@ -794,6 +794,42 @@ jobs:
             eval-conc: ${{ matrix.config['eval-conc'] }}
             scenario-type: agentic-coding
 
+    # Quality-eval rows carry the quality-eval input shape (benchmark-name,
+    # quality-endpoint, quality-model-name, smoke), so they are dispatched
+    # with their own job rather than sweep-evals' fixed-seq-len inputs.
+    sweep-quality-evals:
+        needs: [setup, canary-select, canary-sweep]
+        if: >-
+            ${{
+              !cancelled() &&
+              needs.setup.result == 'success' &&
+              needs.setup.outputs.reuse-enabled != 'true' &&
+              (needs.canary-sweep.result == 'success' || needs.canary-sweep.result == 'skipped') &&
+              toJson(fromJson(needs.setup.outputs.search-space-config).quality_evals) != '[]' &&
+              toJson(fromJson(needs.setup.outputs.search-space-config).quality_evals) != 'null'
+            }}
+        uses: $/.github/workflows/benchmark-tmpl.yml
+        name: quality eval /
+        strategy:
+            fail-fast: ${{ contains(github.event.pull_request.labels.*.name, 'full-sweep-fail-fast') || contains(github.event.pull_request.labels.*.name, 'full-sweep-fail-fast-no-canary') }}
+            matrix:
+                config: ${{ fromJson(needs.setup.outputs.search-space-config).quality_evals }}
+        secrets:
+            REPO_PAT: ${{ secrets.REPO_PAT }}
+            INFERENCEX_OFFICIAL_RO_HF_TOKEN: ${{ secrets.INFERENCEX_OFFICIAL_RO_HF_TOKEN }}
+            GREENNODE_API_KEY: ${{ secrets.GREENNODE_API_KEY }}
+        with:
+            config: ${{ toJSON(matrix.config) }}
+            klaud-run: *klaud-run
+            runner: ${{ matrix.config.runner }}
+            priority: ${{ matrix.config.priority }}
+            queue-token: ${{ matrix.config['queue-token'] }}
+            skip-queue-pr: ${{ matrix.config['skip-queue-pr'] || '' }}
+            run-eval: true
+            eval-only: true
+            eval-limit: ${{ matrix.config.eval-limit || (contains(github.event.pull_request.labels.*.name, 'sweep-enabled') && '10' || '') }}
+            scenario-type: ${{ matrix.config.scenario-type }}
+
     collect-results:
         name: collect-results
         needs:
@@ -825,8 +861,8 @@ jobs:
 
     collect-evals:
         name: collect-evals
-        needs: [sweep-evals, sweep-agentic-evals, sweep-multi-node-evals, sweep-multi-node-agentic-evals, setup]
-        if: ${{ always() && needs.setup.result != 'skipped' && (needs.sweep-evals.result != 'skipped' || needs.sweep-agentic-evals.result != 'skipped' || needs.sweep-multi-node-evals.result != 'skipped' || needs.sweep-multi-node-agentic-evals.result != 'skipped') }}
+        needs: [sweep-evals, sweep-agentic-evals, sweep-multi-node-evals, sweep-multi-node-agentic-evals, sweep-quality-evals, setup]
+        if: ${{ always() && needs.setup.result != 'skipped' && (needs.sweep-evals.result != 'skipped' || needs.sweep-agentic-evals.result != 'skipped' || needs.sweep-multi-node-evals.result != 'skipped' || needs.sweep-multi-node-agentic-evals.result != 'skipped' || needs.sweep-quality-evals.result != 'skipped') }}
         uses: $/.github/workflows/collect-evals.yml
 
     upload-changelog-metadata:

BENCHMARK_TIMES.md

@@ -0,0 +1,23 @@
+# Quality Benchmark Time Estimates
+
+Estimates are for GLM-5.2 at concurrency 4. API latency, reasoning length,
+rate limits, image setup, and sandbox execution can change wall time.
+
+| Benchmark | Full size | Full estimate | 20-sample estimate | Default max generation |
+|---|---:|---:|---:|---:|
+| GPQA-Diamond | 198 | ~20 min | ~2 min | 8K |
+| MMLU-Pro | 12,032 | ~22 h | ~2 min* | 8K |
+| HLE | 2,158 | ~4.5 h | ~5 min* | 16K |
+| LiveCodeBench | 1,055 | ~8 h | ~15 min | 16K |
+| BFCL v4 | ~2,000 | ~1.5 h | <1 min | 8K |
+| SciCode | 288 subproblems | ~5.5 h | ~23 min | 16K |
+| SWE-bench Pro | 731 | ~45 h | ~30 min | 32K per agent turn |
+| DeepSWE | 113 | ~29 h | ~5 h | 32K per agent turn |
+
+\* `LIMIT` applies per subtask for MMLU-Pro and HLE, so `LIMIT=20` evaluates
+more than 20 total questions.
+
+These are output ceilings, not target lengths. Natural stop conditions still
+end generation early. Override any default with `MAX_GEN_TOKENS`; CI matrix
+rows may use `max-gen-tokens`. Streaming avoids idle connection timeouts but
+does not remove total request deadlines or the cost of long generations.