Clean up dataset card: remove slop, add concurrency filtering context

README.md

@@ -344,23 +344,21 @@ LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel CUDA profiles, a
 
 ## Dataset configurations
 
-Five configurations covering serving benchmarks, kernel profiling, and latency predictions.
-
 ### 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.
+Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). 77 (model, hardware, engine) combinations, 17 profiles, 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.
+ISL/OSL sampled from lognormal fits to real workload statistics. 42 combinations, 6 profiles, 7 concurrency levels.
 
 6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
 ### kernels_labeled (148,077 rows)
 
-Per-kernel CUDA profiling data from NCU (Nsight Compute). Individual kernel invocations across 4 GPUs (A100, H100, RTX 3090, RTX 2080Ti) and 13 model/sweep sources. Columns include kernel_family, kernel_name, M/N/K dimensions, gpu_time_duration_ms, dram_bytes_sum, launch_block_size, launch_grid_size, and register pressure.
+Per-kernel CUDA profiles from NCU across 4 GPUs (A100, H100, RTX 3090, RTX 2080 Ti) and 13 model/sweep sources.
 
 ### roofline_quadrant (2,163 rows)
 
@@ -368,11 +366,13 @@ Operational intensity and achieved throughput per kernel, for roofline analysis.
 
 ### predictions (4,715 rows)
 
-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.
+Predicted vs. measured TTFT, TPOT, and E2EL for each serving configuration, with cache-aware prediction metadata. 14 hardware configs.
+
+### Concurrency filtering
 
-### Concurrency levels
+Concurrency is controlled by a fixed-size connection pool. Trace replay uses levels {1, 5, 10, 20, 40, 80}; distributional uses {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}.
+Early runs used a session-pool size smaller than the declared concurrency (`num_sessions=100` for trace replay, `num_sessions=10` for distributional), capping actual load below the nominal value. Rows where declared concurrency exceeded the session pool were dropped.
 
 | Config | Rows |
 |--------|------|
@@ -393,7 +393,7 @@ All benchmarks collected on PyTorch 2.10.0, CUDA 12.8.
 | 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, or 8 GPUs with tensor parallelism. TP degree depends on model size and available GPUs.
+Multi-GPU configurations: 1, 2, 4, or 8 GPUs with tensor parallelism.
 
 ### Models
 
@@ -411,8 +411,6 @@ All models served in BF16 unless noted.
 | gpt-oss-20b | GPT-OSS | 21B (3.6B active) | MoE | mxfp4 projections |
 | gpt-oss-120b | GPT-OSS | 117B (5.1B active) | MoE | mxfp4 projections |
 
-Model names in this table match the `model` column in the parquet files.
-
 ### Engines
 
 - vLLM 0.19.0
@@ -450,7 +448,6 @@ Each row in `summary.parquet` (both configs):
 ```python
 from datasets import load_dataset
 
-# Serving benchmark results
 ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
 # or "distributional", "kernels_labeled", "roofline_quadrant", "predictions"
 ```
@@ -458,33 +455,15 @@ ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
 ## 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.
-- At high concurrency, serving engines may reject requests due to resource limits. Latency and throughput metrics are computed over successful requests only.
+- Metrics: TTFT, TPOT, ITL, E2EL, request throughput, token throughput (mean, median, p90, p99).
+- Metrics 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).
-
-## Planned extensions
-
-- Per-request and multi-turn granularity breakdowns.
-- Additional workload trace corpora.
-
-## Intended uses
-
-- Inference engine comparison under controlled conditions.
-- 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, PyTorch 2.10.0, CUDA 12.8).
-- 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).
+- Distributional profiles are fitted approximations, not direct production replays.
 - Closed-loop concurrency only; no open-loop (Poisson) arrivals.
-- 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