Update README with ISO-Bench scope

README.md

@@ -1,151 +1,62 @@
 ---
 dataset_info:
-  features:
-    - name: item_id
-      dtype: string
-    - name: sample_index
-      dtype: int64
-    - name: agent_name
-      dtype: string
-    - name: model_name
-      dtype: string
-    - name: human_commit
-      dtype: string
-    - name: parent_commit
-      dtype: string
-    - name: benchmark_mode
-      dtype: string
-    - name: perf_command
-      dtype: string
-    - name: llm_model
-      dtype: string
-    - name: status
-      dtype: string
-    - name: error
-      dtype: string
-    - name: benchmark_type
-      dtype: string
-    - name: duration_s
-      dtype: float64
-    - name: ttft_mean_ms
-      dtype: float64
-    - name: ttft_median_ms
-      dtype: float64
-    - name: ttft_p99_ms
-      dtype: float64
-    - name: tpot_mean_ms
-      dtype: float64
-    - name: tpot_median_ms
-      dtype: float64
-    - name: tpot_p99_ms
-      dtype: float64
-    - name: itl_mean_ms
-      dtype: float64
-    - name: itl_median_ms
-      dtype: float64
-    - name: itl_p99_ms
-      dtype: float64
-    - name: request_throughput_req_s
-      dtype: float64
-    - name: output_token_throughput_tok_s
-      dtype: float64
-    - name: total_token_throughput_tok_s
-      dtype: float64
-    - name: latency_avg_ms
-      dtype: float64
-    - name: latency_p50_ms
-      dtype: float64
-    - name: latency_p99_ms
-      dtype: float64
-    - name: throughput_tok_s
-      dtype: float64
-    - name: elapsed_time_s
-      dtype: float64
-    - name: input_throughput_tok_s
-      dtype: float64
-    - name: timestamp
-      dtype: string
   splits:
     - name: train
-      num_examples: 86
+      num_examples: 594
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: data/train-00000-of-00001.parquet
 ---
 
-# Pass@k GPU Benchmark Results
+# ISO-Bench Pass@k GPU Benchmark Results
 
-GPU benchmark results for agent-generated optimization patches from [ISO-Bench](https://huggingface.co/datasets/Lossfunk/ISO-Bench).
+Agent-generated optimization patches benchmarked on real GPU hardware (NVIDIA H100 80GB).
 
-Patches sourced from [Inferencebench/pass-at-k-samples](https://huggingface.co/datasets/Inferencebench/pass-at-k-samples), benchmarked on NVIDIA H100 80GB GPU using Docker-containerized vLLM.
+**Scope:** [Lossfunk/ISO-Bench](https://huggingface.co/datasets/Lossfunk/ISO-Bench) — 54 tasks (39 vLLM + 15 SGLang)
+**Patches from:** [Inferencebench/pass-at-k-samples](https://huggingface.co/datasets/Inferencebench/pass-at-k-samples)
 
 ## Summary
 
-- **86 total rows** (79 successful benchmarks, 7 benchmark failures)
-- **12 tasks** benchmarked across **3 agent/model** configurations
-- Benchmark types: serving, standalone (latency/throughput), prefix caching
-
-## Status Breakdown
-
-| Status | Count | Description |
-|--------|-------|-------------|
-| `success` | 79 | Benchmark completed, metrics captured |
-| `benchmark_failed` | 7 | Patch applied but benchmark errored (server crash, metric parse failure) |
-
-## Agents
-
-| Agent | Model | Successful Benchmarks |
-|-------|-------|-----------------------|
-| claude_code | claude_model-claude-sonnet-4-5 | 69 |
-| claude_code | sonnet | 9 |
-| codex_cli | gpt-5 | 1 |
-
-## Metrics
-
-### Serving Benchmarks
-| Column | Description |
-|--------|-------------|
-| `ttft_mean_ms` | Mean Time To First Token |
-| `ttft_median_ms` | Median Time To First Token |
-| `ttft_p99_ms` | P99 Time To First Token |
-| `tpot_mean_ms` | Mean Time Per Output Token |
-| `tpot_median_ms` | Median Time Per Output Token |
-| `tpot_p99_ms` | P99 Time Per Output Token |
-| `itl_mean_ms` | Mean Inter-Token Latency |
-| `itl_median_ms` | Median Inter-Token Latency |
-| `itl_p99_ms` | P99 Inter-Token Latency |
-| `request_throughput_req_s` | Request throughput (req/s) |
-| `output_token_throughput_tok_s` | Output token throughput (tok/s) |
-| `total_token_throughput_tok_s` | Total token throughput (tok/s) |
-
-### Latency Benchmarks
-| Column | Description |
-|--------|-------------|
-| `latency_avg_ms` | Average latency (ms) |
-| `latency_p50_ms` | P50 latency (ms) |
-| `latency_p99_ms` | P99 latency (ms) |
-| `throughput_tok_s` | Token throughput (tok/s) |
-
-### Prefix Caching Benchmarks
-| Column | Description |
-|--------|-------------|
-| `input_throughput_tok_s` | Input throughput (tok/s) |
-| `throughput_tok_s` | Output throughput (tok/s) |
-| `elapsed_time_s` | Total elapsed time (s) |
-
-## Per-Task Results
-
-| Task | Samples | Benchmark Mode | LLM Model | Avg Throughput |
-|------|---------|---------------|-----------|----------------|
-| vllm_core-0000 | 5 | serving | Qwen/Qwen2.5-7B-Instruct | 3176.4 tok/s |
-| vllm_core-0003 | 7 | serving | deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct | 2124.8 tok/s |
-| vllm_core-0004 | 8 | serving | meta-llama/Meta-Llama-3-8B-Instruct | 2891.3 tok/s |
-| vllm_core-0005 | 8 | serving | meta-llama/Meta-Llama-3-8B-Instruct | 2885.0 tok/s |
-| vllm_core-0006 | 8 | standalone | unknown | 1177.4 tok/s |
-| vllm_core-0007 | 8 | standalone | meta-llama/Meta-Llama-3-8B-Instruct | 8160.0 tok/s |
-| vllm_core-0008 | 8 | serving | meta-llama/Meta-Llama-3-8B-Instruct | 2889.1 tok/s |
-| vllm_core-0009 | 10 | serving | Qwen/Qwen2.5-1.5B-Instruct | 6268.7 tok/s |
-| vllm_core-0010 | 4 | prefix_caching | RedHatAI/Meta-Llama-3-8B-Instruct-FP8 | 5447.6 tok/s |
-| vllm_core-0011 | 3 | prefix_caching | RedHatAI/Meta-Llama-3-8B-Instruct-FP8 | 5380.7 tok/s |
-| vllm_core-0012 | 8 | serving | meta-llama/Meta-Llama-3-8B-Instruct | 2039.1 tok/s |
-| vllm_core-0013 | 2 | serving | meta-llama/Meta-Llama-3-8B-Instruct | 2909.3 tok/s |
+| | vLLM | SGLang | Total |
+|---|---:|---:|---:|
+| **Tasks benchmarked** | **29/39** | **10/15** | **39/54** |
+| **Successful benchmarks** | 444 | 150 | **594** |
+
+| Agent | vLLM | SGLang | Total |
+|-------|-----:|-------:|------:|
+| Claude Code (Sonnet 4.5) | 230 | 76 | 306 |
+| Codex CLI (GPT-5) | 214 | 74 | 288 |
+
+## Schema
+
+| Column | Type | Description |
+|--------|------|-------------|
+| `item_id` | str | ISO-Bench task ID (e.g., `vllm_core-0000`) |
+| `sample_index` | int | Pass@k sample index (0-7) |
+| `repo` | str | `vllm` or `sglang` |
+| `agent_name` | str | `claude_code` or `codex_cli` |
+| `agent_model` | str | `sonnet-4.5` or `gpt-5` |
+| `human_commit` | str | Human optimization commit hash |
+| `parent_commit` | str | Baseline (pre-optimization) commit hash |
+| `benchmark_mode` | str | `serving`, `standalone`, `prefix_caching`, `offline` |
+| `llm_model` | str | LLM model benchmarked (e.g., `meta-llama/Meta-Llama-3-8B-Instruct`) |
+| `duration_s` | float | Total benchmark duration (seconds) |
+
+### Serving Metrics
+`ttft_mean_ms`, `ttft_median_ms`, `ttft_p99_ms` — Time To First Token
+`tpot_mean_ms`, `tpot_median_ms`, `tpot_p99_ms` — Time Per Output Token
+`itl_mean_ms`, `itl_median_ms`, `itl_p99_ms` — Inter-Token Latency
+`request_throughput_req_s` — Requests/sec
+`output_token_throughput_tok_s` — Output tokens/sec
+`total_token_throughput_tok_s` — Total tokens/sec
+
+### Latency Metrics
+`latency_avg_ms`, `latency_p50_ms`, `latency_p99_ms`
+
+### Throughput Metrics
+`throughput_tok_s`, `elapsed_time_s`, `input_throughput_tok_s`
 
 ## Usage
 
@@ -154,9 +65,20 @@ from datasets import load_dataset
 
 ds = load_dataset("Inferencebench/pass-at-k-benchmark-results", split="train")
 
-# Filter to successful benchmarks
-success = ds.filter(lambda x: x["status"] == "success")
+# vLLM Claude Code results
+vllm_cc = ds.filter(lambda x: x["repo"] == "vllm" and x["agent_name"] == "claude_code")
 
-# Get results for a specific task
-task_results = success.filter(lambda x: x["item_id"] == "vllm_core-0004")
+# SGLang Codex results
+sg_cx = ds.filter(lambda x: x["repo"] == "sglang" and x["agent_name"] == "codex_cli")
+
+# Compute pass@k for a specific task
+task = ds.filter(lambda x: x["item_id"] == "vllm_core-0005")
 ```
+
+## Hardware & Method
+
+- **GPU:** NVIDIA H100 80GB PCIe
+- **vLLM:** Docker containers from `shikhar481/vllm_fixed_human_images` (baseline images)
+- **SGLang:** Docker containers from `ayushnangia16/nvidia-sglang-docker`
+- **Execution:** Persistent container per task, sequential samples, setup shared across agents
+- **Date:** 2026-03-29 to 2026-03-31