shikhar007's picture
Update README with ISO-Bench scope
8a62769 verified
metadata
dataset_info:
  splits:
    - name: train
      num_examples: 594
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-00000-of-00001.parquet

ISO-Bench Pass@k GPU Benchmark Results

Agent-generated optimization patches benchmarked on real GPU hardware (NVIDIA H100 80GB).

Scope: Lossfunk/ISO-Bench — 54 tasks (39 vLLM + 15 SGLang) Patches from: Inferencebench/pass-at-k-samples

Summary

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

from datasets import load_dataset

ds = load_dataset("Inferencebench/pass-at-k-benchmark-results", split="train")

# vLLM Claude Code results
vllm_cc = ds.filter(lambda x: x["repo"] == "vllm" and x["agent_name"] == "claude_code")

# 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