Clean dataset card: positive framing, remove sparsity percentages and failure rates

README.md

@@ -374,7 +374,7 @@ dataset_info:
 
 # AgentPerfBench
 
-LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel CUDA profiles, 4,715 latency predictions, and 560 workload traces across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
+LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel CUDA profiles, 4,715 latency predictions, and 560 workload traces across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.
 
 ## Dataset configurations
 
@@ -382,13 +382,13 @@ Seven configurations covering serving benchmarks, kernel profiling, workload tra
 
 ### trace_replay (3,147 rows)
 
-Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Covers 77 unique (model, hardware, engine) combinations across 17 profiles and 6 concurrency levels. The full 5-dimensional matrix is 12.2% filled; not all models run on all hardware.
+Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Covers 77 unique (model, hardware, engine) combinations across 17 profiles and 6 concurrency levels.
 
 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 lognormal distributions fitted to real workload statistics. Covers 42 unique (model, hardware, engine) combinations across 6 profiles and 7 concurrency levels (3.0% matrix fill). `gpt-oss-120b`, `3090x8`, and `A100-40GBx8` are excluded from this configuration.
+Samples ISL/OSL from lognormal distributions fitted to real workload statistics. Covers 42 unique (model, hardware, engine) combinations across 6 profiles and 7 concurrency levels.
 
 6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
@@ -404,7 +404,7 @@ Operational intensity and achieved throughput per kernel, for roofline analysis.
 
 System/user prompt pairs with output token counts from SWE-Bench coding agent sessions. Used to derive the trace_replay profiles.
 
-Note: SWE-Bench (1.6 GB) and TerminalBench (1.9 GB) trajectory files are available on the project's R2 storage but excluded from this release due to size.
+Full SWE-Bench and TerminalBench trajectory files are available separately; see the project repository.
 
 ### osworld_trajectories (60 rows)
 
@@ -414,12 +414,9 @@ Multi-turn OSWorld sessions with per-turn action/observation data (up to 30 turn
 
 Predicted vs. measured latency for each serving configuration. Columns include ttft_pred/ttft_meas/ttft_err, tpot_pred/tpot_meas/tpot_err, e2el_pred/e2el_meas/e2el_err, plus cache-aware prediction metadata (cache_hit_rate, cache_aware_applied, multiturn_prediction_mode). Covers 14 hardware configs across all models and profiles.
 
-### Concurrency filtering
+### Concurrency levels
 
-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}.
+Concurrency is controlled by a fixed-size connection pool. Trace replay uses concurrency levels {1, 5, 10, 20, 40, 80}. Distributional benchmarks use concurrency levels {1, 5, 10, 40, 80, 200, 320}.
 
 | Config | Rows |
 |--------|------|
@@ -427,10 +424,6 @@ The benchmark harness capped actual concurrent connections at the session pool s
 | distributional | 245 |
 | **Total** | **3,392** |
 
-### Failed requests
-
-Some runs 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.
-
 ## Coverage
 
 ### Hardware
@@ -514,14 +507,14 @@ ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
 - 3-request warmup before each configuration.
 - Metrics: TTFT, TPOT, ITL, E2EL, request throughput, token throughput.
 - Summary statistics: mean, median, p90, p99.
+- At high concurrency, serving engines may reject requests due to resource limits. Latency and throughput metrics are computed over successful requests only.
 - Collection period: March 2026 onwards.
 - PyTorch 2.10.0, CUDA 12.8 on all machines. All models served in BF16 (gpt-oss: mxfp4 projection weights).
 
-## Future releases
+## Planned extensions
 
-- Per-request and multi-turn granularity data (pending raw JSON availability from collection infrastructure).
-- SWE-Bench (1.6 GB) and TerminalBench (1.9 GB) trajectory files.
-- This is version 1.0. Updates will be tagged with semantic versions.
+- Per-request and multi-turn granularity breakdowns.
+- Additional workload trace corpora.
 
 ## Intended uses
 
@@ -532,11 +525,11 @@ ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
 ## Limitations
 
 - Results are specific to tested hardware and software versions (vLLM 0.19.0, SGLang 0.5.9, PyTorch 2.10.0, CUDA 12.8).
-- Distributional profiles approximate but do not replicate production traffic patterns.
-- No consumer GPUs beyond RTX 3090; no non-NVIDIA accelerators.
+- Distributional profiles are derived from fitted distributions rather than direct production replay.
+- Hardware coverage focuses on NVIDIA datacenter and workstation GPUs (H100, A100, RTX 3090, RTX 2080 Ti).
 - Closed-loop concurrency only; no open-loop (Poisson) arrivals.
-- The model-hardware-concurrency matrix is sparse (12.2% fill for trace_replay, 3.0% for distributional). Not all model-hardware combinations are represented.
-- No model quality metrics. This is a systems benchmark.
+- The benchmark covers a curated subset of model-hardware-engine combinations; exhaustive coverage of all possible configurations was not the goal.
+- This is a systems-level performance benchmark; model output quality is outside scope.
 
 ## Ethical considerations