Feature Request: Publish Calibration Dataset for Dynamic-GGUF-Quants-2.0

[rikunarita]

Feature Request: Publish Calibration Dataset for Dynamic-GGUF-Quants-2.0

Summary

I would like to request the release of the calibration dataset used for Dynamic-GGUF-Quants-2.0, so that users can reproduce Unsloth's quantization quality with their own pipelines.

Background

Unsloth's Dynamic 2.0 quantization achieves excellent results with "new imatrix data" (300K-1.5M+ tokens) that is hand-curated and cleaned. However, the imatrix_unsloth.gguf_file included in GGUF repositories is a pre-computed importance matrix, not the raw calibration text data.

Problem

• The imatrix_unsloth.gguf_file cannot be reverse-engineered to recover the original calibration text.
• Users who want to quantize models with custom settings (different bit-widths, mixed-precision schemes, or non-standard tensor types) cannot leverage Unsloth's high-quality calibration data.
• The quantization recipes are tied to Unsloth's internal pipeline, limiting reproducibility for advanced users.

Request

Please consider releasing one of the following:

1. Raw calibration text dataset (or a script to generate it)

   • Even a partial/subset release would be valuable
   • Could be hosted on Hugging Face as a separate dataset repository

2. Documentation on dataset composition

   • What domains/topics are covered (code, math, dialogue, etc.)
   • Approximate token counts and filtering criteria

3. Standard-format imatrix export

   • If the current imatrix_unsloth.gguf_file is already llama.cpp-compatible, please document this explicitly
   • If not, consider providing a converter or native export option

Use Cases

• Researchers reproducing Unsloth's quantization results in academic papers
• Users creating custom GGUF variants (e.g., specialized medical/legal models)
• Community members improving quantization pipelines for niche hardware

Related

• Blog post: https://unsloth.ai/blog/dynamic-gguf
• Dynamic 2.0 announcement mentions "hand curated and cleaned" calibration data

Thank you for considering this request!