datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
GLM-5.3-Flash-BF16-Teacher-Logits
GLM-5.3-Flash BF16 teacher logits
This dataset contains full-vocabulary float32 teacher logits from the immutable
zai-org/GLM-5.3-Flash-BF16 revision a6c167b62691b2bac901344b65cb651a70f53e43.
It keeps the sealed final KLD panel qualification-only and publishes the
separate non-final calibration panel under role-specific paths.
Qualification-only final windows: 25
Qualification-only final prediction positions: 51175
Vocabulary size: 154880
Teacher receipt:… See the full description on the dataset page: https://huggingface.co/datasets/brandonmusic/GLM-5.3-Flash-BF16-Teacher-Logits.qwen3.6-27b-h100-bf16-benchmark
Two GPUs do not mean twice the users
This study answers a serving decision, not a hardware trivia question: when a
27B model already fits on one H100, should a second GPU shard the model or run a
second independent replica?
The answer
Concurrency is a load-generator setting, not a user count and not a promise.
Capacity is the highest real arrival rate that satisfies a declared service
objective. That is why this study measures both a saturated concurrency curve… See the full description on the dataset page: https://huggingface.co/datasets/cogeanu-marius/qwen3.6-27b-h100-bf16-benchmark.opentq-qwen36-bf16-sidecar
Qwen3.6-27B BF16 Sidecar Runs
This dataset publishes BF16 sidecar outputs used to compare OpenTQ GGUF artifacts against the base model Qwen/Qwen3.6-27B on pinned practical mini-subsets.
The Dataset Viewer uses flattened Parquet tables so columns have stable types. The original raw JSON files remain in runs/<job_id>/ for reproducibility.
Tables
Config
Grain
Purpose
results
one row per benchmark sample
prompts, task IDs, deterministic BF16 outputs, score fields… See the full description on the dataset page: https://huggingface.co/datasets/zlaabsi/opentq-qwen36-bf16-sidecar.