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`](https://huggingface.co/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, task IDs, scoring metadata and BF16/Q3/Q4/Q5 outputs. | | `data/paired_summary.jsonl` | One row per benchmark plus `TOTAL`, matching the model card table. | | `raw/bf16/no_think.json` | BF16 sidecar output from Hugging Face Jobs H200 run `69f235d2d2c8bd8662bd320e` on `Qwen/Qwen3.6-27B`. | | `raw/quant/q3.json` | Local stock `llama.cpp`/Metal/FlashAttention run for `Q3_K_M`. | | `raw/quant/q4.json` | Local stock `llama.cpp`/Metal/FlashAttention run for `Q4_K_M`. | | `raw/quant/q5.json` | Local stock `llama.cpp`/Metal/FlashAttention run for `Q5_K_M`. | | `reports/paired_summary.csv` | Benchmark summary CSV. | | `reports/paired_summary.json` | Benchmark summary JSON. | | `reports/paired_report.md` | Human-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.