CoolFace
Datasetpublic

zlaabsi/Qwen3.6-27B-OTQ-GGUF-benchmarks

Qwen3.6-27B OTQ GGUF Benchmark Reproducibility This dataset contains the compact paired benchmark evidence used by zlaabsi/Qwen3.6-27B-OTQ-GGUF. It is a reproducibility dataset, not a leaderboard dataset. The rows are small practical release signals run on pinned task IDs with prompt format qwen3-no-think, deterministic decoding and local scoring rules. Contents Path Meaning data/paired_samples.jsonl Flattened 232-row paired sample table with prompts… See the full description on the dataset page: https://huggingface.co/datasets/zlaabsi/Qwen3.6-27B-OTQ-GGUF-benchmarks.

sourceHugging Faceotherupdated 5mo agoView on Hugging Face
1likes14downloads
Dataset Card

Qwen3.6-27B OTQ GGUF Benchmark Reproducibility

This dataset contains the compact paired benchmark evidence used by `zlaabsi/Qwen3.6-27B-OTQ-GGUF`.

It is a reproducibility dataset, not a leaderboard dataset. The rows are small practical release signals run on pinned task IDs with prompt format qwen3-no-think, deterministic decoding and local scoring rules.

Contents

PathMeaning
data/paired_samples.jsonlFlattened 232-row paired sample table with prompts, task IDs, scoring metadata and BF16/Q3/Q4/Q5 outputs.
data/paired_summary.jsonlOne row per benchmark plus TOTAL, matching the model card table.
raw/bf16/no_think.jsonBF16 sidecar output from Hugging Face Jobs H200 run 69f235d2d2c8bd8662bd320e on Qwen/Qwen3.6-27B.
raw/quant/q3.jsonLocal stock llama.cpp/Metal/FlashAttention run for Q3_K_M.
raw/quant/q4.jsonLocal stock llama.cpp/Metal/FlashAttention run for Q4_K_M.
raw/quant/q5.jsonLocal stock llama.cpp/Metal/FlashAttention run for Q5_K_M.
reports/paired_summary.csvBenchmark summary CSV.
reports/paired_summary.jsonBenchmark summary JSON.
reports/paired_report.mdHuman-readable paired subset report.

Scope

  • —Base model: Qwen/Qwen3.6-27B
  • —GGUF model repo: zlaabsi/Qwen3.6-27B-OTQ-GGUF
  • —BF16 runtime: Hugging Face Jobs H200 with Transformers
  • —GGUF runtime: stock llama.cpp / llama-server, Metal, FlashAttention
  • —Prompt format: qwen3-no-think
  • —Sample count: 232

These files include prompts derived from upstream benchmark datasets. Respect the upstream dataset licenses and terms for any reuse beyond reproducibility of this release evaluation.