CoolFace
Datasetpublic

ssakethch/h200-quantization-benchmarks

H200 Quantization Benchmarks Benchmark results for 40 quantized and non-quantized instruction-tuned LLMs evaluated on an NVIDIA H200 MIG (Multi-Instance GPU) setup. This dataset supports reproducible comparison of quantization methods (AWQ, GPTQ, fp8, bf16) across accuracy and throughput dimensions. Dataset Configs Config Description Rows accuracy Per-task accuracy results from lm-eval across all models ~240 accuracy_leaderboard Aggregated accuracy… See the full description on the dataset page: https://huggingface.co/datasets/ssakethch/h200-quantization-benchmarks.

sourceHugging Faceupdated 5mo agoView on Hugging Face
1likes17downloads
Dataset Card

H200 Quantization Benchmarks

Benchmark results for 40 quantized and non-quantized instruction-tuned LLMs evaluated on an NVIDIA H200 MIG (Multi-Instance GPU) setup. This dataset supports reproducible comparison of quantization methods (AWQ, GPTQ, fp8, bf16) across accuracy and throughput dimensions.

Dataset Configs

ConfigDescriptionRows
accuracyPer-task accuracy results from lm-eval across all models~240
accuracy_leaderboardAggregated accuracy scores per model for leaderboard ranking
model_inventoryRegistry of all evaluated models with quantization metadata
throughputToken throughput measurements (tokens/sec) via vLLM
throughput_leaderboardAggregated throughput scores per model for leaderboard ranking

Hardware

  • GPU: NVIDIA H200 SXM (141 GB HBM3e)
  • Setup: MIG (Multi-Instance GPU) partitioned instances
  • Inference engine: vLLM

Quantization Methods

MethodDescription
awqActivation-aware Weight Quantization (INT4)
gptqGPTQ post-training quantization (INT4)
fp88-bit floating point (native H200 support)
bf16BFloat16 baseline (no quantization)

Evaluation Tasks

Accuracy benchmarks are run with lm-evaluation-harness across:

  • arc_challenge — ARC Challenge (25-shot)
  • hellaswag — HellaSwag (10-shot)
  • winogrande — Winogrande (5-shot)
  • gsm8k — GSM8K math reasoning (5-shot)
  • mmlu / mmlu_abstract_algebra — MMLU (5-shot)

Usage

python
from datasets import load_dataset

# Load accuracy benchmarks
accuracy = load_dataset("ssakethch/h200-quantization-benchmarks", "accuracy")

# Load model inventory
inventory = load_dataset("ssakethch/h200-quantization-benchmarks", "model_inventory")

# Load throughput benchmarks
throughput = load_dataset("ssakethch/h200-quantization-benchmarks", "throughput")

# Load leaderboards
acc_lb = load_dataset("ssakethch/h200-quantization-benchmarks", "accuracy_leaderboard")
thr_lb = load_dataset("ssakethch/h200-quantization-benchmarks", "throughput_leaderboard")

Schema

accuracy

ColumnTypeDescription
timestampstringUnix timestamp of the evaluation run
modelstringFull HuggingFace model ID
model_familystringModel family (e.g., Llama-3.2, Qwen3)
model_sizestringParameter count (e.g., 3B, 7B, 14B)
quantstringQuantization method
tierstringEvaluation tier (tier1 / tier2)
taskstringBenchmark task name
n_shotfloatNumber of few-shot examples
metricstringMetric used (e.g., acc, acc_norm)
accuracyfloatTask accuracy score
stderrfloatStandard error of accuracy
samples_originalfloatOriginal number of samples
samples_effectivefloatEffective samples after filtering

model_inventory

ColumnTypeDescription
modelstringFull HuggingFace model ID
model_familystringModel family
model_sizestringParameter count
quantstringQuantization method
has_accuracyboolWhether accuracy benchmarks exist
has_throughputboolWhether throughput benchmarks exist

License

MIT