512 samples form the exact calibration blend used to GPTQ-quantize the canada-quant/hy3-w4a16-mtp model. The dataset was published by canada-quant on Hugging Face for full reproducibility of the quantization pipeline. It was last updated on July 17, 2026.
Use Cases
- Reproducing the GPTQ quantization pipeline for the Hy3 model based on the described calibration set.
- Benchmarking quantization performance on code and tool-call-shaped tokens as mentioned in the description.
- Analyzing the effect of blended calibration sets versus chat-only sets on routed expert activation.
- Studying calibration strategies for W4A16 weight quantization of large language models.
Strengths
- Provides the exact 512-sample calibration set used for a specific model quantization, ensuring reproducibility.
- Deliberately constructed as a blend to address under-sampling of tokens that activate routed experts, as described.
- Published with a clear purpose for full pipeline transparency on a major platform (Hugging Face).
Limitations
- Description metadata is limited; actual data quality and content require manual inspection after download.
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is known (512), but the specific content and format of the samples are not detailed.
Provenance
- Source
- canada-quant on Hugging Face.
- Collection Method
- Created as a calibration blend for GPTQ quantization of the tencent/Hy3 model.
- Freshness
- Last updated 2026-07-17 18:13:16; freshness should be verified.