Upload folder using huggingface_hub

README.md

@@ -10,20 +10,19 @@ tags:
   - gpu-profiling
   - vllm
   - sglang
-  - roofline
   - agentic-workloads
 size_categories:
   - 1K<n<10K
 pretty_name: AgentPerfBench
 configs:
-  - config_name: benchmark_results
+  - config_name: trace_replay
     data_files:
       - split: summary
-        path: benchmark_results/summary.parquet
-      - split: per_request
-        path: benchmark_results/per_request.parquet
-      - split: multi_turn
-        path: benchmark_results/multi_turn.parquet
+        path: trace_replay/summary.parquet
+  - config_name: distributional
+    data_files:
+      - split: summary
+        path: distributional/summary.parquet
   - config_name: roofline
     data_files:
       - split: kernel_profiles
@@ -32,75 +31,103 @@ configs:
 
 # AgentPerfBench
 
-## Dataset Summary
+LLM inference benchmark dataset measuring serving performance (TTFT, TPOT, ITL, throughput) across 9 models, 14 GPU configurations, 2 serving engines, and 20+ workload profiles spanning single-turn chat, multi-turn agent sessions, and synthetic stress tests. Includes per-kernel CUDA profiling data for roofline analysis.
 
-AgentPerfBench measures LLM inference serving performance under agentic workloads. The dataset contains benchmark results (TTFT, TPOT, ITL, throughput) across 9 models, 3 GPU platforms, 2 serving engines, and 14+ workload profiles spanning single-turn chat, multi-turn agent sessions, and synthetic stress tests. It also includes per-kernel CUDA profiling data for roofline analysis. Data was collected in April 2026.
+## Dataset Configurations
 
-## Data Description
+This dataset provides two benchmark configurations reflecting distinct data collection methodologies:
 
-### Parquet Files
+### trace_replay
 
-| File | Description | Rows (approx.) |
-|------|-------------|-----------------|
-| `benchmark_results/summary.parquet` | One row per benchmark configuration (model x hardware x engine x profile x concurrency). Columns: `run_id`, `model`, `hardware`, `engine`, `tensor_parallelism`, `profile`, `concurrency`, plus TTFT/TPOT/ITL/E2EL percentiles and throughput metrics. | 2500-3500 |
-| `benchmark_results/per_request.parquet` | Per-request latency data. FK: `run_id` references `summary`. | 150K-200K |
-| `benchmark_results/multi_turn.parquet` | Per-turn breakdowns for multi-turn sessions. FK: `run_id` references `summary`. | varies |
-| `roofline/kernel_profiles.parquet` | Per-kernel CUDA profiling data from Nsight Compute. One row per kernel invocation. | varies |
+Requests replay exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Input distributions are empirically grounded in real tool-use patterns, capturing realistic burstiness and turn-depth correlations.
 
-### Metadata JSONs
+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`
 
-- `metadata/models.json` — model specifications (family, parameters, architecture, license)
-- `metadata/hardware.json` — GPU specs (VRAM, bandwidth, peak TFLOPS)
-- `metadata/profiles.json` — workload profile definitions (ISL/OSL ranges, source, tier)
+### distributional
 
-## Benchmark Methodology
+Requests sample ISL/OSL from parameterized statistical distributions (e.g., lognormal) fitted to real workload statistics. Provides controlled experimental variation while matching aggregate characteristics.
 
-- **Concurrency model**: Closed-loop with semaphore control. One request admitted per completion.
-- **Concurrency sweep**: 1 to 320.
-- **Requests per configuration**: 50, with 3-request warmup.
-- **Metrics**: TTFT, TPOT (time per output token), ITL (inter-token latency), E2EL (end-to-end latency), request throughput (req/s), token throughput (tok/s).
-- **Percentiles**: mean, median, p90, p99.
-- **Kernel profiling**: PyTorch profiler on 2-layer forward passes, batch sizes [1, 4, 8, 32, 64].
+Profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
+
+### Why two configurations?
+
+**trace_replay** proves relevance: patterns are drawn from real agent sessions, grounding results in observed behavior. **distributional** proves mechanism: controlled statistical sampling isolates the effects of workload parameters from incidental correlations. Together they support both ecological validity and experimental control.
 
-## Models
+### Concurrency filtering
 
-| Model | Family | Parameters | Architecture | License |
-|-------|--------|-----------|--------------|---------|
-| Llama-3.1-8B-Instruct | Llama | 8B | Dense | Llama 3.1 Community |
-| Llama-3.1-70B-Instruct | Llama | 70B | Dense | Llama 3.1 Community |
-| Llama-3.3-70B-Instruct | Llama | 70B | Dense | Llama 3.3 Community |
-| Qwen2.5-72B-Instruct | Qwen | 72B | Dense | Apache 2.0 |
-| Qwen3.5-9B | Qwen | 9B | Dense | Apache 2.0 |
-| Qwen3.5-27B | Qwen | 27B | Dense | Apache 2.0 |
-| gpt-oss-20b | GPT-OSS | 21B (3.6B active) | MoE (32 experts, top-4) | Apache 2.0 |
-| gpt-oss-120b | GPT-OSS | 117B (5.1B active) | MoE (128 experts, top-4) | Apache 2.0 |
+Rows where declared concurrency exceeds the session pool size have been excluded. This affects trace_replay data at concurrency > 100 (session pool was 100) and distributional/current data at concurrency > 10 (session pool was 10). Distributional data collected after the fix has no such limitation.
 
-## Hardware
+## Coverage
+
+### Hardware
 
 | GPU | VRAM | HBM Bandwidth | Peak BF16 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.
+
+### 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 |
+| 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 |
+
+### Engines
+
+- vLLM 0.19.0
+- SGLang 0.5.9
+
+## Schema
+
+Each row in `summary.parquet` (both configs) contains:
+
+| Column | Type | Description |
+|--------|------|-------------|
+| run_id | string | Deterministic hash of run parameters |
+| model | string | Model short name |
+| model_family | string | Model family (llama, qwen, gpt-oss, mixtral) |
+| hardware | string | GPU configuration (e.g., H100x4) |
+| engine | string | Serving engine (vllm, sglang) |
+| tensor_parallelism | int | TP degree |
+| profile | string | Workload profile name |
+| concurrency | int | Concurrent request level |
+| num_requests | int | Total requests in run |
+| duration_s | float | Total run duration |
+| successful_requests | int | Completed requests |
+| failed_requests | int | Failed requests |
+| request_throughput | float | Requests/second |
+| input_token_throughput | float | Input tokens/second |
+| output_token_throughput | float | Output tokens/second |
+| total_token_throughput | float | Total tokens/second |
+| mean/median/p90/p99_ttft_ms | float | Time to first token |
+| mean/median/p90/p99_tpot_ms | float | Time per output token |
+| mean/median/p90/p99_itl_ms | float | Inter-token latency |
+| mean/median/p90/p99_e2el_ms | float | End-to-end latency |
+
+## Benchmark Methodology
+
+- **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].
+
+## Future Releases
 
-## Workload Profiles
-
-| Profile | Tier | ISL | OSL | Source | Synthetic? |
-|---------|------|-----|-----|--------|------------|
-| chat-short | 2 (Chat) | 500 | 300 | ShareGPT | No |
-| chat-medium | 2 (Chat) | 2000 | 1000 | ShareGPT | No |
-| chat-long | 2 (Chat) | 8000 | 2000 | ShareGPT | No |
-| coding-agent | 1 (Agentic) | 17000 | 800 | SWE-Bench | No |
-| chat-multiturn-short | 2 (Chat MT) | 8192 | 1000 | ShareGPT | No |
-| chat-multiturn-medium | 2 (Chat MT) | 16384 | 1500 | ShareGPT | No |
-| chat-multiturn-long | 2 (Chat MT) | 32768 | 2000 | ShareGPT | No |
-| swebench-multiturn-short | 1 (Agentic MT) | 32768 | 2000 | SWE-Bench | No |
-| swebench-multiturn-medium | 1 (Agentic MT) | 65536 | 2000 | SWE-Bench | No |
-| terminalbench-multiturn-short | 1 (Agentic MT) | 32768 | 2000 | TerminalBench | No |
-| terminalbench-multiturn-medium | 1 (Agentic MT) | 65536 | 2000 | TerminalBench | No |
-| decode-heavy | 3 (Synthetic) | 256 | 4096 | Random | Yes |
-| prefill-heavy | 3 (Synthetic) | 8192 | 256 | Random | Yes |
-| random-1k | 3 (Synthetic) | 1024 | 1024 | Random | Yes |
+Per-request and multi-turn granularity data will be added when full result JSON files are available from the collection infrastructure.
 
 ## Intended Uses
 
@@ -112,7 +139,7 @@ AgentPerfBench measures LLM inference serving performance under agentic workload
 ## Limitations
 
 - Results are specific to tested hardware and software versions (vLLM 0.19.0, SGLang 0.5.9).
-- Synthetic profiles approximate but do not replicate production traffic patterns.
+- Distributional profiles approximate but do not replicate exact production traffic patterns.
 - No consumer GPUs beyond RTX 3090; no non-NVIDIA accelerators.
 - Closed-loop concurrency only; open-loop (Poisson arrival) not included.
 - No model quality metrics. This is a systems benchmark.
@@ -120,15 +147,15 @@ AgentPerfBench measures LLM inference serving performance under agentic workload
 ## Ethical Considerations
 
 - No PII in the dataset.
-- Synthetic profiles use random tokens. Real-trace profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench).
+- Synthetic profiles use random tokens. Trace-replay profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench, OSWorld).
 - Benchmark results should not be used as sole basis for hardware purchasing decisions.
 
 ## Source Datasets
 
-- [SWE-Bench](https://github.com/princeton-nlp/SWE-bench) (MIT License) — coding agent traces
-- [TerminalBench](https://github.com/TerminalBench/TerminalBench) — terminal agent traces
-- [ShareGPT (Aeala/ShareGPT_Vicuna_unfiltered)](https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered) — chat baseline prompts (used as input length distribution, not redistributed)
-- [OSWorld](https://github.com/xlang-ai/OSWorld) — computer-use agent traces
+- [SWE-Bench](https://github.com/princeton-nlp/SWE-bench) (MIT License)
+- [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)
 
 ## Citation
 

benchmark_results/multi_turn.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b357a1e1a233ab00e0c57e39fd7c7c0be07e4d3c3f1ec3419ed43757afff4a8b
+size 1831697

benchmark_results/per_request.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd4b87d4365983bed1184d5ae024e2d0056f53e45d183f4dd178f673340fc3d5
+size 15337735

benchmark_results/summary.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0123d2758fd1aee6161b36786e5b879f0bae5cf96dad1e51a79fe64035ec9fbd
+size 540814

croissant.json

@@ -49,12 +49,12 @@
   },
   "@type": "sc:Dataset",
   "name": "AgentPerfBench",
-  "description": "LLM inference benchmark dataset measuring serving performance (TTFT, TPOT, throughput) across 9 models, 3 GPU platforms, 2 serving engines under agentic and chat workloads.",
+  "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",
   "license": "https://spdx.org/licenses/Apache-2.0.html",
   "conformsTo": "http://mlcommons.org/croissant/1.1",
-  "datePublished": "2026-05-03",
-  "version": "1.0.0",
+  "datePublished": "2026-05-04",
+  "version": "2.0.0",
   "citeAs": "@inproceedings{agentperfbench2026, title={AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs}, author={Anonymous}, booktitle={NeurIPS 2026 Evaluations and Datasets Track}, year={2026}}",
   "creator": {
     "@type": "sc:Organization",
@@ -63,32 +63,24 @@
   "distribution": [
     {
       "@type": "cr:FileObject",
-      "@id": "summary-parquet",
-      "name": "summary.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/benchmark_results/summary.parquet",
+      "@id": "trace-replay-summary-parquet",
+      "name": "trace_replay/summary.parquet",
+      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/trace_replay/summary.parquet",
       "encodingFormat": "application/x-parquet",
-      "sha256": "b59eea237b64b0abe21b9d981e1d5dbb935a970333a091dbeca53a4f9f9c7a5b"
+      "sha256": "a9cd2565a9292e75f54791e8108f2491b7bb371fdfeeefb81cdb00833860dd23"
     },
     {
       "@type": "cr:FileObject",
-      "@id": "per-request-parquet",
-      "name": "per_request.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/benchmark_results/per_request.parquet",
+      "@id": "distributional-summary-parquet",
+      "name": "distributional/summary.parquet",
+      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/distributional/summary.parquet",
       "encodingFormat": "application/x-parquet",
-      "sha256": "fd4b87d4365983bed1184d5ae024e2d0056f53e45d183f4dd178f673340fc3d5"
-    },
-    {
-      "@type": "cr:FileObject",
-      "@id": "multi-turn-parquet",
-      "name": "multi_turn.parquet",
-      "contentUrl": "https://huggingface.co/datasets/lynae-1219/AgentPerfBench/resolve/main/benchmark_results/multi_turn.parquet",
-      "encodingFormat": "application/x-parquet",
-      "sha256": "b357a1e1a233ab00e0c57e39fd7c7c0be07e4d3c3f1ec3419ed43757afff4a8b"
+      "sha256": "b97f8709d2cc87725697b26e2dbc2c3f6a5db52b7883ea6888e5626c9be25d7b"
     },
     {
       "@type": "cr:FileObject",
       "@id": "kernel-profiles-parquet",
-      "name": "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"
@@ -97,18 +89,18 @@
   "recordSet": [
     {
       "@type": "cr:RecordSet",
-      "@id": "benchmark-summary",
-      "name": "Benchmark Summary",
-      "description": "One row per benchmark configuration with aggregated serving metrics.",
+      "@id": "trace-replay-summary",
+      "name": "Trace Replay Summary",
+      "description": "One row per benchmark configuration from trace replay runs (empirical agent session replays).",
       "field": [
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/run_id",
+          "@id": "trace-replay-summary/run_id",
           "name": "run_id",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "run_id"
@@ -117,12 +109,12 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/model",
+          "@id": "trace-replay-summary/model",
           "name": "model",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "model"
@@ -131,12 +123,12 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/hardware",
+          "@id": "trace-replay-summary/hardware",
           "name": "hardware",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "hardware"
@@ -145,12 +137,12 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/engine",
+          "@id": "trace-replay-summary/engine",
           "name": "engine",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "engine"
@@ -159,12 +151,12 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/profile",
+          "@id": "trace-replay-summary/profile",
           "name": "profile",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "profile"
@@ -173,12 +165,12 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/concurrency",
+          "@id": "trace-replay-summary/concurrency",
           "name": "concurrency",
           "dataType": "sc:Integer",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
               "column": "concurrency"
@@ -187,43 +179,43 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/median_ttft_ms",
-          "name": "median_ttft_ms",
+          "@id": "trace-replay-summary/request_throughput",
+          "name": "request_throughput",
           "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
-              "column": "median_ttft_ms"
+              "column": "request_throughput"
             }
           }
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/median_tpot_ms",
-          "name": "median_tpot_ms",
+          "@id": "trace-replay-summary/median_ttft_ms",
+          "name": "median_ttft_ms",
           "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
-              "column": "median_tpot_ms"
+              "column": "median_ttft_ms"
             }
           }
         },
         {
           "@type": "cr:Field",
-          "@id": "benchmark-summary/request_throughput",
-          "name": "request_throughput",
+          "@id": "trace-replay-summary/median_tpot_ms",
+          "name": "median_tpot_ms",
           "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "summary-parquet"
+              "@id": "trace-replay-summary-parquet"
             },
             "extract": {
-              "column": "request_throughput"
+              "column": "median_tpot_ms"
             }
           }
         }
@@ -231,18 +223,18 @@
     },
     {
       "@type": "cr:RecordSet",
-      "@id": "per-request-data",
-      "name": "Per-Request Latency",
-      "description": "Individual request latency measurements. Foreign key: run_id references benchmark-summary.",
+      "@id": "distributional-summary",
+      "name": "Distributional Summary",
+      "description": "One row per benchmark configuration from distributional runs (statistical sampling from fitted distributions).",
       "field": [
         {
           "@type": "cr:Field",
-          "@id": "per-request-data/run_id",
+          "@id": "distributional-summary/run_id",
           "name": "run_id",
           "dataType": "sc:Text",
           "source": {
             "fileObject": {
-              "@id": "per-request-parquet"
+              "@id": "distributional-summary-parquet"
             },
             "extract": {
               "column": "run_id"
@@ -251,57 +243,113 @@
         },
         {
           "@type": "cr:Field",
-          "@id": "per-request-data/request_idx",
-          "name": "request_idx",
+          "@id": "distributional-summary/model",
+          "name": "model",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "distributional-summary-parquet"
+            },
+            "extract": {
+              "column": "model"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "distributional-summary/hardware",
+          "name": "hardware",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "distributional-summary-parquet"
+            },
+            "extract": {
+              "column": "hardware"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "distributional-summary/engine",
+          "name": "engine",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "distributional-summary-parquet"
+            },
+            "extract": {
+              "column": "engine"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "distributional-summary/profile",
+          "name": "profile",
+          "dataType": "sc:Text",
+          "source": {
+            "fileObject": {
+              "@id": "distributional-summary-parquet"
+            },
+            "extract": {
+              "column": "profile"
+            }
+          }
+        },
+        {
+          "@type": "cr:Field",
+          "@id": "distributional-summary/concurrency",
+          "name": "concurrency",
           "dataType": "sc:Integer",
           "source": {
             "fileObject": {
-              "@id": "per-request-parquet"
+              "@id": "distributional-summary-parquet"
             },
             "extract": {
-              "column": "request_idx"
+              "column": "concurrency"
             }
           }
         },
         {
           "@type": "cr:Field",
-          "@id": "per-request-data/ttft_ms",
-          "name": "ttft_ms",
+          "@id": "distributional-summary/request_throughput",
+          "name": "request_throughput",
           "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "per-request-parquet"
+              "@id": "distributional-summary-parquet"
             },
             "extract": {
-              "column": "ttft_ms"
+              "column": "request_throughput"
             }
           }
         },
         {
           "@type": "cr:Field",
-          "@id": "per-request-data/tpot_ms",
-          "name": "tpot_ms",
+          "@id": "distributional-summary/median_ttft_ms",
+          "name": "median_ttft_ms",
           "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "per-request-parquet"
+              "@id": "distributional-summary-parquet"
             },
             "extract": {
-              "column": "tpot_ms"
+              "column": "median_ttft_ms"
             }
           }
         },
         {
           "@type": "cr:Field",
-          "@id": "per-request-data/success",
-          "name": "success",
-          "dataType": "sc:Boolean",
+          "@id": "distributional-summary/median_tpot_ms",
+          "name": "median_tpot_ms",
+          "dataType": "sc:Float",
           "source": {
             "fileObject": {
-              "@id": "per-request-parquet"
+              "@id": "distributional-summary-parquet"
             },
             "extract": {
-              "column": "success"
+              "column": "median_tpot_ms"
             }
           }
         }
@@ -386,11 +434,11 @@
       ]
     }
   ],
-  "rai:dataLimitations": "Results cover NVIDIA H100, A100, and RTX 3090 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 datacenter GPUs. Workload profiles are author-designed approximations of production traffic; real deployment patterns may differ. ShareGPT chat baselines reflect English-language conversations only.",
-  "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. Raw user conversations from ShareGPT are not included in this dataset.",
+  "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:dataSocialImpact": "Enables reproducible comparison of open-source LLM serving systems, supporting infrastructure research and reducing vendor lock-in. Could be misused for misleading marketing claims about model or hardware performance without proper context.",
+  "rai:dataSocialImpact": "Enables reproducible comparison of open-source LLM serving systems, supporting infrastructure research and reducing vendor lock-in.",
   "rai:hasSyntheticData": true,
   "prov:wasDerivedFrom": [
     {
@@ -400,11 +448,15 @@
     {
       "@id": "https://github.com/TerminalBench/TerminalBench",
       "name": "TerminalBench"
+    },
+    {
+      "@id": "https://github.com/xlang-ai/OSWorld",
+      "name": "OSWorld"
     }
   ],
   "prov:wasGeneratedBy": {
     "@type": "prov:Activity",
     "name": "AgentPerfBench benchmark collection",
-    "description": "Deploy model on target GPU with specified engine and tensor parallelism. Send 50 requests per configuration after 3-request warmup using closed-loop concurrency control. Record per-request TTFT, TPOT, ITL, E2EL, and token counts. Compute summary percentiles. For roofline data: profile 2 transformer layers using PyTorch profiler. 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 percentiles. Sanitize credentials and convert to Parquet."
   }
-}
+}

distributional/summary.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b97f8709d2cc87725697b26e2dbc2c3f6a5db52b7883ea6888e5626c9be25d7b
+size 50325

trace_replay/summary.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9cd2565a9292e75f54791e8108f2491b7bb371fdfeeefb81cdb00833860dd23
+size 694254