Split multi-schema configs into individual configs (5 -> 7) for datasets lib compat

README.md

@@ -24,21 +24,25 @@ configs:
     data_files:
       - split: summary
         path: distributional/summary.parquet
-  - config_name: kernel_profiles
+  - config_name: kernels_labeled
     data_files:
-      - split: kernels_labeled
+      - split: train
         path: kernel_profiles/kernels_labeled.parquet
-      - split: roofline_quadrant
+  - config_name: roofline_quadrant
+    data_files:
+      - split: train
         path: kernel_profiles/roofline_quadrant.parquet
-  - config_name: workload_traces
+  - config_name: coding_agent_prompts
     data_files:
-      - split: coding_agent_prompts
+      - split: train
         path: workload_traces/coding_agent_prompts.parquet
-      - split: osworld_trajectories
+  - config_name: osworld_trajectories
+    data_files:
+      - split: train
         path: workload_traces/osworld_trajectories.parquet
   - config_name: predictions
     data_files:
-      - split: serving_predictions
+      - split: train
         path: predictions/serving_predictions.parquet
 dataset_info:
   - config_name: trace_replay
@@ -181,17 +185,105 @@ dataset_info:
       - name: summary
         num_examples: 245
         num_bytes: 70836
-  - config_name: kernel_profiles
+  - config_name: kernels_labeled
+    features:
+      - name: source
+        dtype: string
+      - name: gpu
+        dtype: string
+      - name: model
+        dtype: string
+      - name: kernel_family
+        dtype: string
+      - name: kernel_name
+        dtype: string
+      - name: dtype
+        dtype: string
+      - name: held_out
+        dtype: bool
+      - name: M
+        dtype: float64
+      - name: N
+        dtype: float64
+      - name: K
+        dtype: float64
+      - name: bs
+        dtype: float64
+      - name: seq
+        dtype: float64
+      - name: n_heads
+        dtype: float64
+      - name: head_dim
+        dtype: float64
+      - name: kv_heads
+        dtype: float64
+      - name: numel
+        dtype: float64
+      - name: op_type
+        dtype: string
+      - name: gpu_time_duration_ms
+        dtype: float64
+      - name: launch_block_size
+        dtype: float64
+      - name: launch_grid_size
+        dtype: float64
+      - name: dram_bytes_sum
+        dtype: float64
+      - name: launch_registers_per_thread
+        dtype: float64
     splits:
-      - name: kernels_labeled
+      - name: train
         num_examples: 148077
-      - name: roofline_quadrant
+  - config_name: roofline_quadrant
+    features:
+      - name: model
+        dtype: string
+      - name: profile
+        dtype: string
+      - name: concurrency
+        dtype: int64
+      - name: engine
+        dtype: string
+      - name: hardware
+        dtype: string
+      - name: oi
+        dtype: float64
+      - name: cf_gb
+        dtype: float64
+      - name: output_tput
+        dtype: float64
+      - name: tpot_ms
+        dtype: float64
+      - name: ttft_ms
+        dtype: float64
+    splits:
+      - name: train
         num_examples: 2163
-  - config_name: workload_traces
+  - config_name: coding_agent_prompts
+    features:
+      - name: system
+        dtype: string
+      - name: user
+        dtype: string
+      - name: osl_tokens
+        dtype: int64
     splits:
-      - name: coding_agent_prompts
+      - name: train
         num_examples: 500
-      - name: osworld_trajectories
+  - config_name: osworld_trajectories
+    features:
+      - name: session_id
+        dtype: string
+      - name: source
+        dtype: string
+      - name: num_turns
+        dtype: int64
+      - name: turns
+        dtype: string
+      - name: run
+        dtype: string
+    splits:
+      - name: train
         num_examples: 60
   - config_name: predictions
     features:
@@ -276,7 +368,7 @@ dataset_info:
       - name: measurement_semantics_warning
         dtype: string
     splits:
-      - name: serving_predictions
+      - name: train
         num_examples: 4715
 ---
 
@@ -286,7 +378,7 @@ LLM inference benchmark: 3,392 serving runs, 148,077 per-kernel CUDA profiles, 4
 
 ## Dataset configurations
 
-Five configurations covering serving benchmarks, kernel profiling, workload traces, and latency predictions.
+Seven configurations covering serving benchmarks, kernel profiling, workload traces, and latency predictions.
 
 ### trace_replay (3,147 rows)
 
@@ -300,22 +392,24 @@ Samples ISL/OSL from lognormal distributions fitted to real workload statistics.
 
 6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn`
 
-### kernel_profiles (148,077 + 2,163 rows)
+### kernels_labeled (148,077 rows)
 
-Per-kernel CUDA profiling data from NCU (Nsight Compute). Two splits:
+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.
 
-- **kernels_labeled** (148,077 rows): 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.
-- **roofline_quadrant** (2,163 rows): Operational intensity and achieved throughput per kernel, for roofline analysis. H100 reference hardware (989 peak TFLOPS, 3.35 TB/s HBM).
+### roofline_quadrant (2,163 rows)
 
-### workload_traces (500 + 60 rows)
+Operational intensity and achieved throughput per kernel, for roofline analysis. H100 reference hardware (989 peak TFLOPS, 3.35 TB/s HBM).
 
-Raw agent session recordings used to derive the trace_replay profiles.
+### coding_agent_prompts (500 rows)
 
-- **coding_agent_prompts** (500 rows): System/user prompt pairs with output token counts from SWE-Bench coding agent sessions.
-- **osworld_trajectories** (60 rows): Multi-turn OSWorld sessions with per-turn action/observation data (up to 30 turns per session).
+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.
 
+### osworld_trajectories (60 rows)
+
+Multi-turn OSWorld sessions with per-turn action/observation data (up to 30 turns per session). Used to derive the trace_replay profiles.
+
 ### 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.
@@ -409,7 +503,8 @@ from datasets import load_dataset
 
 # Serving benchmark results
 ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
-# or "distributional", "kernel_profiles", "workload_traces", "predictions"
+# or "distributional", "kernels_labeled", "roofline_quadrant",
+# "coding_agent_prompts", "osworld_trajectories", "predictions"
 ```
 
 ## Benchmark methodology