Fix dataset card: accurate stats, failed request disclosure, croissant alignment, BF16 column fix

README.md

@@ -23,39 +23,40 @@ configs:
     data_files:
       - split: summary
         path: distributional/summary.parquet
-  - config_name: roofline
-    data_files:
-      - split: kernel_profiles
-        path: roofline/kernel_profiles.parquet
 ---
 
 # AgentPerfBench
 
-LLM inference benchmark dataset: serving performance (TTFT, TPOT, ITL, throughput) across 9 models, 14 GPU configurations, 2 engines, and 20+ workload profiles covering single-turn chat, multi-turn agent sessions, and stress tests. Includes per-kernel CUDA profiling for roofline analysis.
+LLM inference benchmark measuring TTFT, TPOT, ITL, and throughput across 9 models, up to 14 GPU configurations, 2 engines, and 21 workload profiles.
 
-## Dataset Configurations
+## Dataset configurations
 
-Two benchmark configurations, each with a different data collection method.
+Two configs with different data collection methods.
 
 ### trace_replay (3,147 rows)
 
-Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Input distributions come from real tool-use patterns.
+Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT).
 
-Profiles: `chat-medium`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `decode-heavy`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `prefill-heavy`, `random-1k`, `swebench-multiturn-medium`, `swebench-multiturn-short`, `terminalbench-multiturn-medium`, `terminalbench-multiturn-short`
+17 profiles: `chat-medium`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `decode-heavy`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `prefill-heavy`, `random-1k`, `swebench-multiturn-medium`, `swebench-multiturn-short`, `terminalbench-multiturn-medium`, `terminalbench-multiturn-short`
 
 ### distributional (245 rows)
 
-Samples ISL/OSL from parameterized distributions (e.g., lognormal) fitted to real workload statistics. Faster to run than full trace replays; MSE-validated against trace_replay.
+Samples ISL/OSL from parameterized distributions (lognormal) fitted to real workload statistics. Validated against trace_replay.
+
+8 of 9 models covered (`gpt-oss-120b` excluded). 12 of 14 hardware configs (`3090x8`, `A100-40GBx8` excluded).
 
-Profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
+6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
 ### Why two configurations?
 
-trace_replay grounds results in observed agent behaviour; distributional covers the model-hardware-concurrency matrix faster.
+trace_replay uses exact sequences from recorded sessions; distributional samples from fitted distributions for broader coverage with shorter runs.
 
 ### Concurrency filtering
 
-The benchmark harness capped actual concurrent connections at the session pool size. Rows where declared concurrency exceeds the pool size never achieved the intended load and have been excluded. This affects trace_replay data at concurrency > 100 (session pool was 100) and distributional data collected before the fix at concurrency > 10 (session pool was 10). Distributional data collected after the fix has no such limitation.
+The benchmark harness capped actual concurrent connections at the session pool size. Rows where declared concurrency exceeded the pool were excluded.
+
+- trace_replay: concurrency > 100 removed (session pool was 100). Remaining values: {1, 5, 10, 20, 40, 80}.
+- distributional (pre-fix): concurrency > 10 removed (session pool was 10). Post-fix data has no cap. Remaining values: {1, 5, 10, 40, 80, 200, 320}.
 
 | Config | Rows |
 |--------|------|
@@ -63,33 +64,39 @@ The benchmark harness capped actual concurrent connections at the session pool s
 | distributional | 245 |
 | **Total** | **3,392** |
 
+### Failed requests
+
+Some configurations produce request failures, typically at high concurrency where the engine hits memory or timeout limits. 30.8% of trace_replay rows and 42% of distributional rows have `failed_requests > 0`. Summary metrics (TTFT, TPOT, throughput) are computed from successful requests only. The `failed_requests` column is included for transparency.
+
 ## Coverage
 
 ### Hardware
 
-| GPU | VRAM | HBM Bandwidth | Peak BF16 TFLOPS |
+| GPU | VRAM | HBM bandwidth | Peak half-precision TFLOPS |
 |-----|------|---------------|------------------|
 | NVIDIA H100 SXM | 80 GB | 3.35 TB/s | 989 |
 | NVIDIA A100 SXM4 | 40 GB | 1.56 TB/s | 312 |
 | NVIDIA RTX 3090 | 24 GB | 936 GB/s | 71 |
 | NVIDIA RTX 2080 Ti | 11 GB | 616 GB/s | 27 |
 
-Multi-GPU configurations: 1, 2, 4, 8 GPUs with tensor parallelism.
+Multi-GPU configurations: 1, 2, 4, or 8 GPUs with tensor parallelism (TP degree depends on GPU and model).
 
 ### Models
 
 | Model | Family | Parameters | Architecture |
 |-------|--------|-----------|--------------|
-| Llama-3.1-8B-Instruct | Llama | 8B | Dense |
-| Llama-3.1-70B-Instruct | Llama | 70B | Dense |
-| Llama-3.3-70B-Instruct | Llama | 70B | Dense |
-| Qwen2.5-72B-Instruct | Qwen | 72B | Dense |
+| Llama-3.1-8B | Llama | 8B | Dense |
+| Llama-3.1-70B | Llama | 70B | Dense |
+| Llama-3.3-70B | Llama | 70B | Dense |
+| Qwen2.5-72B | Qwen | 72B | Dense |
 | Qwen3.5-9B | Qwen | 9B | Dense |
 | Qwen3.5-27B | Qwen | 27B | Dense |
 | Mixtral-8x7B | Mixtral | 46.7B (12.9B active) | MoE |
 | gpt-oss-20b | GPT-OSS | 21B (3.6B active) | MoE |
 | gpt-oss-120b | GPT-OSS | 117B (5.1B active) | MoE |
 
+Model names in this table match the `model` column in the parquet files.
+
 ### Engines
 
 - vLLM 0.19.0
@@ -97,7 +104,7 @@ Multi-GPU configurations: 1, 2, 4, 8 GPUs with tensor parallelism.
 
 ## Schema
 
-Each row in `summary.parquet` (both configs) contains:
+Each row in `summary.parquet` (both configs):
 
 | Column | Type | Description |
 |--------|------|-------------|
@@ -122,41 +129,53 @@ Each row in `summary.parquet` (both configs) contains:
 | mean/median/p90/p99_itl_ms | float | Inter-token latency |
 | mean/median/p90/p99_e2el_ms | float | End-to-end latency |
 
-## Benchmark Methodology
+## Loading
 
-- **Concurrency model**: Closed-loop with semaphore control.
-- **Concurrency sweep**: 1 to 320.
-- **Requests per configuration**: 50-100, with 3-request warmup.
-- **Metrics**: TTFT, TPOT, ITL, E2EL, request throughput, token throughput.
-- **Percentiles**: mean, median, p90, p99.
-- **Kernel profiling** (roofline config): PyTorch profiler on 2-layer forward passes, batch sizes [1, 4, 8, 32, 64].
+```python
+from datasets import load_dataset
 
-## Future Releases
+ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
+# or "distributional"
+```
+
+## Benchmark methodology
+
+- Closed-loop concurrency with semaphore control.
+- Concurrency levels: {1, 5, 10, 20, 40, 80} (trace_replay), {1, 5, 10, 40, 80, 200, 320} (distributional).
+- 3-request warmup before each configuration.
+- Metrics: TTFT, TPOT, ITL, E2EL, request throughput, token throughput.
+- Summary statistics: mean, median, p90, p99.
 
-Per-request and multi-turn granularity data will be added when full result JSON files are available from the collection infrastructure.
+## Future releases
 
-## Intended Uses
+- Per-request and multi-turn granularity data (pending raw JSON availability).
+- Per-kernel CUDA roofline profiles (PyTorch profiler, 2-layer forward passes, batch sizes 1/4/8/32/64).
+
+## Intended uses
 
 - Inference engine comparison under controlled conditions.
-- Capacity planning for agentic LLM deployments.
-- GPU roofline analysis across workload types.
-- Studying TTFT degradation under multi-turn context accumulation.
+- Capacity planning for LLM deployments.
+- TTFT scaling with context length in multi-turn sessions.
 
 ## Limitations
 
 - Results are specific to tested hardware and software versions (vLLM 0.19.0, SGLang 0.5.9).
-- Distributional profiles approximate but do not replicate exact production traffic patterns.
+- Distributional profiles approximate but do not replicate production traffic patterns.
 - No consumer GPUs beyond RTX 3090; no non-NVIDIA accelerators.
-- Closed-loop concurrency only; no open-loop (Poisson arrival).
+- Closed-loop concurrency only; no open-loop (Poisson) arrivals.
 - No model quality metrics. This is a systems benchmark.
 
-## Ethical Considerations
+## Ethical considerations
+
+No PII. Trace-replay profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench, OSWorld). Synthetic profiles use random tokens.
+
+## License
 
-No PII. Synthetic profiles use random tokens; trace-replay profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench, OSWorld). Not a basis for hardware purchasing decisions on its own.
+The benchmark data is released under Apache-2.0. Source datasets retain their original licenses (see below).
 
-## Source Datasets
+## Source datasets
 
-- [SWE-Bench](https://github.com/princeton-nlp/SWE-bench) (MIT License)
+- [SWE-Bench](https://github.com/princeton-nlp/SWE-bench) (MIT)
 - [TerminalBench](https://github.com/TerminalBench/TerminalBench)
 - [ShareGPT (Aeala/ShareGPT_Vicuna_unfiltered)](https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered)
 - [OSWorld](https://github.com/xlang-ai/OSWorld)

croissant.json

@@ -50,7 +50,7 @@
   "@type": "sc:Dataset",
   "name": "AgentPerfBench",
   "description": "LLM inference benchmark dataset measuring serving performance (TTFT, TPOT, throughput) across 9 models, 4 GPU platforms, 2 serving engines under agentic and chat workloads. Provides two dataset configurations: trace_replay (empirical session replays) and distributional (statistical sampling).",
-  "url": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench",
+  "url": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench",
   "license": "https://spdx.org/licenses/Apache-2.0.html",
   "conformsTo": "http://mlcommons.org/croissant/1.1",
   "datePublished": "2026-05-04",
@@ -65,7 +65,7 @@
       "@type": "cr:FileObject",
       "@id": "trace-replay-summary-parquet",
       "name": "trace_replay/summary.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/trace_replay/summary.parquet",
+      "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/trace_replay/summary.parquet",
       "encodingFormat": "application/x-parquet",
       "sha256": "a9cd2565a9292e75f54791e8108f2491b7bb371fdfeeefb81cdb00833860dd23"
     },
@@ -73,17 +73,9 @@
       "@type": "cr:FileObject",
       "@id": "distributional-summary-parquet",
       "name": "distributional/summary.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/distributional/summary.parquet",
+      "contentUrl": "https://huggingface.co/datasets/agent-perf-bench/AgentPerfBench/resolve/main/distributional/summary.parquet",
       "encodingFormat": "application/x-parquet",
       "sha256": "bb9a48351e0b612b3125c5d0a8e200638d790269f821e45aae342a4042bfd951"
-    },
-    {
-      "@type": "cr:FileObject",
-      "@id": "kernel-profiles-parquet",
-      "name": "roofline/kernel_profiles.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/roofline/kernel_profiles.parquet",
-      "encodingFormat": "application/x-parquet",
-      "sha256": "109cf3206b6cef47ad033f93db2e943a124c0a5a55e9db7bdf2bdbc9b864290a"
     }
   ],
   "recordSet": [
@@ -225,7 +217,7 @@
       "@type": "cr:RecordSet",
       "@id": "distributional-summary",
       "name": "Distributional Summary",
-      "description": "One row per benchmark configuration from distributional runs (statistical sampling, shorter to run, MSE-validated against trace replays).",
+      "description": "One row per benchmark configuration from distributional runs (statistical sampling, validated against trace_replay).",
       "field": [
         {
           "@type": "cr:Field",
@@ -354,90 +346,12 @@
           }
         }
       ]
-    },
-    {
-      "@type": "cr:RecordSet",
-      "@id": "kernel-profile-data",
-      "name": "Kernel Profiles",
-      "description": "Per-kernel CUDA profiling data for roofline analysis.",
-      "field": [
-        {
-          "@type": "cr:Field",
-          "@id": "kernel-profile-data/model",
-          "name": "model",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "kernel-profiles-parquet"
-            },
-            "extract": {
-              "column": "model"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "kernel-profile-data/phase",
-          "name": "phase",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "kernel-profiles-parquet"
-            },
-            "extract": {
-              "column": "phase"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "kernel-profile-data/kernel_name",
-          "name": "kernel_name",
-          "dataType": "sc:Text",
-          "source": {
-            "fileObject": {
-              "@id": "kernel-profiles-parquet"
-            },
-            "extract": {
-              "column": "kernel_name"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "kernel-profile-data/arithmetic_intensity",
-          "name": "arithmetic_intensity",
-          "dataType": "sc:Float",
-          "source": {
-            "fileObject": {
-              "@id": "kernel-profiles-parquet"
-            },
-            "extract": {
-              "column": "arithmetic_intensity"
-            }
-          }
-        },
-        {
-          "@type": "cr:Field",
-          "@id": "kernel-profile-data/achieved_tflops",
-          "name": "achieved_tflops",
-          "dataType": "sc:Float",
-          "source": {
-            "fileObject": {
-              "@id": "kernel-profiles-parquet"
-            },
-            "extract": {
-              "column": "achieved_tflops"
-            }
-          }
-        }
-      ]
     }
   ],
   "rai:dataLimitations": "Results cover NVIDIA H100, A100, RTX 3090, and RTX 2080 Ti GPUs only and may not generalize to other accelerators (AMD, Intel, TPU). Benchmark configurations are pinned to vLLM 0.19.0 and SGLang 0.5.9; results do not represent other engine versions. Concurrency levels (1-320) may not cover extreme-scale deployments. Not recommended as sole basis for hardware purchasing decisions or for comparing model task quality.",
   "rai:dataBiases": "Model selection over-represents Meta Llama and Alibaba Qwen families. Hardware is exclusively NVIDIA GPUs. Workload profiles are author-designed approximations of production traffic; real deployment patterns may differ.",
   "rai:personalSensitiveInformation": "No personally identifiable information is present. All API endpoints and credentials are stripped. Workload traces use synthetic random tokens or publicly available coding benchmarks.",
-  "rai:dataUseCases": "Established uses: relative comparison of inference engine throughput, latency benchmarking under controlled conditions, GPU roofline analysis, studying TTFT degradation under multi-turn context growth. Not established: absolute latency prediction for production, model quality comparison, cost estimation.",
+  "rai:dataUseCases": "Established uses: relative comparison of inference engine throughput, latency benchmarking under controlled conditions, GPU roofline analysis, TTFT scaling with context length in multi-turn sessions. Not established: absolute latency prediction for production, model quality comparison, cost estimation.",
   "rai:dataSocialImpact": "Enables reproducible comparison of open-source LLM serving systems, supporting infrastructure research and reducing vendor lock-in.",
   "rai:hasSyntheticData": true,
   "prov:wasDerivedFrom": [
@@ -457,6 +371,6 @@
   "prov:wasGeneratedBy": {
     "@type": "prov:Activity",
     "name": "AgentPerfBench benchmark collection",
-    "description": "Deploy model on target GPU with specified engine and tensor parallelism. Send requests per configuration after warmup using closed-loop concurrency control. Record per-request TTFT, TPOT, ITL, E2EL, and token counts. Compute summary percentiles. Sanitize credentials and convert to Parquet."
+    "description": "Deploy model on target GPU with specified engine and tensor parallelism. Send requests per configuration after warmup using closed-loop concurrency control. Record per-request TTFT, TPOT, ITL, E2EL, and token counts. Compute summary statistics. Sanitize credentials and convert to Parquet."
   }
 }