Clean up dataset card: simplify retention table, remove AI-generated prose

README.md

@@ -31,38 +31,37 @@ configs:
 
 # AgentPerfBench
 
-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.
+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.
 
 ## Dataset Configurations
 
-This dataset provides two benchmark configurations reflecting distinct data collection methodologies:
+Two benchmark configurations, each with a different data collection method.
 
 ### trace_replay (3,147 rows)
 
-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.
+Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Input distributions come from real tool-use patterns.
 
 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)
 
-Requests sample ISL/OSL from parameterized statistical distributions (e.g., lognormal) fitted to real workload statistics. Shorter to run than full trace replays, enabling faster characterization of profile-level serving behavior across the hardware matrix. MSE validation confirms distributional runs reproduce the latency and throughput characteristics of their trace-replay counterparts.
+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.
 
 Profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
 ### Why two configurations?
 
-**trace_replay** provides ecological validity — patterns are drawn from real agent sessions, grounding results in observed behavior. **distributional** enables efficient coverage — shorter run times allow systematic sweeps across the full model-hardware-concurrency matrix, while MSE validation against trace_replay confirms the results remain representative.
+trace_replay grounds results in observed agent behaviour; distributional covers the model-hardware-concurrency matrix faster.
 
 ### 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.
+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.
 
-| Scope | Raw rows | After filter | Removed | Filter rule |
-|-------|----------|-------------|---------|-------------|
-| trace_replay (archive) | 4,814 | 3,147 | 1,667 | C > 100 (session pool = 100) |
-| distributional (current) | 843 | 159 | 684 | C > 10 (session pool = 10) |
-| distributional (fixed) | 86 | 86 | 0 | No cap — bug fixed |
-| **Total** | **5,743** | **3,392** | **2,351** | **Retention: 59.1%** |
+| Config | Rows |
+|--------|------|
+| trace_replay | 3,147 |
+| distributional | 245 |
+| **Total** | **3,392** |
 
 ## Coverage
 
@@ -138,9 +137,9 @@ Per-request and multi-turn granularity data will be added when full result JSON
 
 ## Intended Uses
 
-- Comparing inference engine performance under controlled conditions.
+- Inference engine comparison under controlled conditions.
 - Capacity planning for agentic LLM deployments.
-- Roofline analysis of GPU utilization under different workload regimes.
+- GPU roofline analysis across workload types.
 - Studying TTFT degradation under multi-turn context accumulation.
 
 ## Limitations
@@ -148,14 +147,12 @@ Per-request and multi-turn granularity data will be added when full result JSON
 - 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.
 - No consumer GPUs beyond RTX 3090; no non-NVIDIA accelerators.
-- Closed-loop concurrency only; open-loop (Poisson arrival) not included.
+- Closed-loop concurrency only; no open-loop (Poisson arrival).
 - No model quality metrics. This is a systems benchmark.
 
 ## Ethical Considerations
 
-- No PII in the dataset.
-- 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.
+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.
 
 ## Source Datasets