ahmedBargady/open-models-benchmark-results

⚡ Local LLM Evaluation Leaderboard

Welcome to the official public benchmark leaderboard maintained by @ahmedBargady.
This dataset repository hosts benchmark evaluation metrics, accuracy scores, throughput telemetry, and quantization trade-off analyses of open-weights foundation models tested locally on NVIDIA A100 GPUs.

---

💻 Hardware & System Specifications

All evaluations are executed under standardized local cluster environments:

Specification	Details
Maintainer	@ahmedBargady
GPU Setup	2x NVIDIA A100-SXM4-80GB
Total VRAM	160 GB
Inference Engine	vLLM (v0.27.1) / CUDA 13.2
Evaluation Framework	EleutherAI lm-evaluation-harness

---

📐 Benchmark Metrics & What The Scores Mean

To provide a rigorous evaluation of each model, we evaluate across three primary dimensions: Mathematical Reasoning, Commonsense NLI, and Real-World Prefill Throughput.

1. 🧮 GSM8K (Grade School Math 8K) — Mathematical & Step-by-Step Reasoning

• What it measures: A dataset of 8,500 high-quality linguistically diverse grade school math word problems requiring multi-step reasoning, arithmetic calculations, and chain-of-thought logic.
• Evaluation Metric: Strict Exact Match (exact_match,strict-match / exact_match,none). The model must produce the exact numerical answer following its chain of reasoning.
• What the score means:
  • > 85.0%: Superior mathematical and multi-step reasoning capabilities (typical of distilled reasoning models like DeepSeek-R1-Distill or optimized instruction models).
  • 70.0% - 85.0%: Strong general reasoning capability; handles complex multi-step word problems well.
  • < 60.0%: Struggles with multi-step arithmetic accuracy or strict formatting constraints.

2. 🧠 HellaSwag — Commonsense Reasoning & Natural Language Inference

• What it measures: A benchmark designed to test a model's commonsense inference by asking it to choose the most plausible continuation for a given real-world scenario (e.g., physical activities, cooking, daily human interactions).
• Evaluation Metric: Acc Norm (acc_norm,none — Length-Normalized Accuracy). Normalizes log-likelihood scores by option length to prevent bias toward shorter responses.
• What the score means:
  • > 82.0%: State-of-the-art commonsense comprehension and physical world context alignment.
  • 75.0% - 82.0%: Solid natural language comprehension and context completion.
  • ~ 25.0%: Random guessing baseline (4-choice multiple choice).

3. ⚡ Input Speed (tok/s) — Prefill Throughput & Inference Telemetry

• What it measures: The prefill processing speed in tokens per second (est. speed input) measured live during batch evaluation under vLLM.
• Why it matters: Higher input throughput means faster prompt processing, lower Time-To-First-Token (TTFT), and better compute efficiency for high-concurrency production deployments.
• Key Hardware Scaling Observations:
  • TP=2 (Dual A100): Reaches ~6,200+ tok/s via GPU-to-GPU NVLink tensor parallelism.
  • TP=1 (Single A100): Ranges between 2,000 - 4,200 tok/s depending on model architecture, attention implementation, and quantization precision (bfloat16, FP8, AWQ).

---

🏆 Interactive Leaderboard Table

Rendered directly from data/train.csv via Hugging Face's Dataset Viewer.

Model Name	Hardware / Execution	Precision	Input Speed (tok/s)	GSM8K (Strict Exact Match)	HellaSwag (Acc Norm)	Date
Qwen/Qwen3.8-27B	2x NVIDIA A100-SXM4-80GB (TP=2)	bfloat16	6,213.6	70.36%	82.85%	2026-08-22
Qwen/Qwen3.8-27B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	2,724.9	70.36%	82.80%	2026-08-22
Qwen/Qwen3.8-27B-FP8	1x NVIDIA A100-SXM4-80GB (TP=1)	fp8	2,063.3	68.76%	82.91%	2026-08-22
deepseek-ai/DeepSeek-R1-Distill-Qwen-32B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	4,114.2	82.41%	81.86%	2026-08-22
casperhansen/deepseek-r1-distill-qwen-32b-awq	1x NVIDIA A100-SXM4-80GB (TP=1)	awq-int4	3,036.5	86.20%	80.89%	2026-08-23
stelterlab/Mistral-Small-24B-Instruct-2501-AWQ	1x NVIDIA A100-SXM4-80GB (TP=1)	awq-int4	4,240.3	89.16%	82.98%	2026-08-23
orcarouter/Qwen3.8-27B-Uncensored-FP8	1x NVIDIA A100-SXM4-80GB (TP=1)	fp8	2,890.7	73.16%	82.91%	2026-08-23
RedHatAI/DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic	1x NVIDIA A100-SXM4-80GB (TP=1)	fp8-dynamic	3,571.9	81.73%	81.86%	2026-08-23
ornith-ai/Ornith-1.5-35B-A3B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	5,982.0	50.27%	81.02%	2026-08-23
mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	3,895.0	58.76%	82.45%	2026-08-23
Qwen/Qwen3-8B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	16,348.7	87.57%	74.89%	2026-08-23
Qwen/Qwen3-0.6B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	134,702.1	42.00%	47.22%	2026-08-23
deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B	1x NVIDIA A100-SXM4-80GB (TP=1)	bfloat16	63,708.2	70.36%	44.80%	2026-08-23
nvidia/Qwen3.6-35B-A3B-NVFP4	1x NVIDIA A100-SXM4-80GB (TP=1)	nvfp4	7,782.5	2.43%	83.02%	2026-08-23

---

🔬 Key Takeaways & Trade-off Analysis

1. Quantization Impact (FP8 vs BF16):
   • GSM8K Math Reasoning: Full bfloat16 achieves 70.36%, while FP8 quantized version scores 68.76% (a modest ~1.6% drop for 50% lower VRAM footprint).
   • HellaSwag Commonsense NLI: Both models exhibit near-identical performance (82.80% vs 82.91%).
2. Tensor Parallel Scaling (2x A100 vs 1x A100):
   • Scaling across 2x A100 GPUs via Tensor Parallelism increases prefill prompt throughput from 2,724.9 tok/s to 6,213.6 tok/s (2.28x speedup).

---

🛠️ How to Reproduce Local Evaluations

1. Dual-GPU Full Precision (bfloat16) Execution (TP=2)

source .venv-vllm/bin/activate
export HF_HOME=/srv/lustre01/project/jepa-qdy1am7soje/users/ahmed.bargady/open-weights-models/.cache/huggingface
export TMPDIR=/tmp
export VLLM_USE_FLASHINFER_SAMPLER=0
export FLASHINFER_DISABLE_JIT=1

lm_eval --model vllm \
  --model_args pretrained=Qwen/Qwen3.8-27B,dtype=bfloat16,tensor_parallel_size=2,gpu_memory_utilization=0.9,trust_remote_code=True \
  --tasks gsm8k,hellaswag \
  --batch_size auto \
  --output_path hf_bench_results.json \
  --log_samples

2. Single-GPU Quantized (FP8) Execution (GPU 1)

source .venv-vllm/bin/activate
export HF_HOME=/srv/lustre01/project/jepa-qdy1am7soje/users/ahmed.bargady/open-weights-models/.cache/huggingface
export TMPDIR=/tmp
export CUDA_VISIBLE_DEVICES=1

lm_eval --model vllm \
  --model_args pretrained=Qwen/Qwen3.8-27B-FP8,dtype=bfloat16,tensor_parallel_size=1,max_model_len=4096,gpu_memory_utilization=0.85,trust_remote_code=True \
  --tasks gsm8k,hellaswag \
  --batch_size auto \
  --output_path hf_bench_results_single_gpu_fp8.json \
  --log_samples

---

📂 Repository File Structure

• data/train.csv: Structured CSV tabular backend driving HF's native Dataset Viewer.
• leaderboard.csv: Master copy of benchmark summary metrics.
• hf_bench_results.json: Raw output JSON containing per-sample prompt logs, likelihoods, and model configs.
• README.md: Dataset Card documentation.