RemySkye/bartowski_v5_semantic_25p

Bartowski v5 Semantic — 25% compact calibration set

I made this for faster llama.cpp iMatrix generation while keeping the result close to the full dataset.

Why I made this

I wanted a quicker way to build iMatrices. The full Bartowski v5 semantic set is good, but it takes longer than I need when I am testing many models and quants. I started with the original dataset by lemon07r and kept a coverage-aware 25% subset instead of taking only the first rows.

My goal was simple: save a lot of calibration time while keeping quality close to the full dataset. This is a speed-focused option, not a claim that fewer examples can never lose quality.

Credit

The original dataset and semantic calibration work are by lemon07r:

lemon07r/bartowski-imatrix-v5-semantic

This upload is only a deterministic, smaller selection from that dataset. I am not claiming the original data as my own. The source is marked MIT on Hugging Face, and I keep the attribution here so the credit stays visible.

What is inside

I selected 519 of the 2,075 source rows (25%). I kept examples from different parts of the source, content types, scripts, and length ranges. The text was not rewritten, translated, or generated.

Content type	Rows
Code	187
Prose	186
Math	71
Prompt	43
Document	25
Dialogue	7
Total	519

The set is mainly English, but it also includes Cyrillic, Arabic, Greek, Devanagari, Chinese/Japanese, and other Latin-script examples.

Quick facts

	Full source	This dataset
Rows	2,075	519
File size	1,496,006 bytes	396,700 bytes
Estimated Qwen3.5 tokens	about 356,630	about 92,098
iMatrix chunks at context 512	696	179

That is 517 fewer chunks and about 74% less model-evaluation work. In the calibration part of the run, that was roughly 4× faster. The exact wall-clock time depends on the model and hardware; larger models usually save more time.

What I measured

I compared IQ4_XS iMatrices made with the full source and this 25% subset on Qwen3.5-0.8B. Both runs used stock CUDA llama.cpp, context 512, special-token parsing, and the same 32 WikiText-2 evaluation chunks.

Metric	Full source	25% subset	Change
Calibration rows	2,075	519	75.0% fewer
Calibration chunks	696	179	74.3% fewer
IQ4_XS perplexity	18.541455	18.593929	+0.28%
Mean KL divergence	0.052578	0.052492	-0.16%
Matrix cosine similarity	1.000000	0.998979	99.90%

The numbers are close enough for my fast-iteration use case. Your result can still vary with another model, quantization type, or evaluation set.

Why I chose 25%

I tested several sizes against the full source. Smaller can be faster, but the margin becomes less dependable. The 25% point gave me a useful balance.

Subset	Rows	IQ4_XS PPL	Relative KL change	My read
100%	2,075	18.541455	+0.00%	Reference
50%	1,038	18.614800	-0.45%	Good
35%	726	18.525937	-0.57%	Good
25%	519	18.593929	-0.16%	Best speed/safety balance
20%	415	18.562700	-0.89%	Still close
15%	311	18.642213	-0.95%	Fast, less safe
10%	208	18.821507	+1.67%	Too risky for my default

Small cross-model check

I also used the 15% test as a quick stress check on other architectures. This is not a replacement for testing the 25% file on your own model.

Model	Full IQ4_XS PPL	15% PPL	Relative KL change
Qwen3.5-0.8B	18.541455	18.642213	-0.95%
Gemma 4 E2B	139.294382	143.925528	+0.12%
LFM2.5-230M	45.596583	44.419854	+0.44%

Treat these as checks, not guarantees. For a production quant or a very small quant, I would still compare against a full-data iMatrix first.

Audit summary

I checked the source rows, selected indices, packaged text, and saved iMatrix.

• All 2,075 source records parsed successfully.
• The 519 selected indices are unique, sorted, and in range.
• Every selected row is non-empty and at least 200 characters long.
• No exact or whitespace-normalized duplicate rows were found.
• The text is valid UTF-8 with no BOM, NUL bytes, replacement characters, or stray control characters.
• The saved Qwen3.5 iMatrix reports chunk_count = 179 and chunk_size = 512. There are four valid zero-width-space characters inherited from the source. I left them unchanged rather than silently changing the benchmarked text.

Files in this upload

• bartowski_v5_semantic_25p.txt — the calibration text used by llama-imatrix.
• selection_manifest.json — machine-readable provenance with the source, selected row indices, coverage counts, and hashes. It is optional metadata; it is not another calibration file.

Generate an iMatrix with llama.cpp

Build

sudo apt-get update
sudo apt-get install -y build-essential cmake git
git clone https://github.com/ggml-org/llama.cpp.git
cmake -S llama.cpp -B llama.cpp/build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build llama.cpp/build --config Release --target llama-imatrix -j2

For a CPU-only build, remove -DGGML_CUDA=ON.

Generate the file

Put this text file and your BF16 model in the same working folder:

./llama.cpp/build/bin/llama-imatrix \
  -m Qwen3.5-0.8B-BF16.gguf \
  -f bartowski_v5_semantic_25p.txt \
  -o bartowski_v5_semantic_25p.imatrix.gguf \
  --ctx-size 512 \
  --parse-special \
  --no-ppl \
  -ngl 99 -b 512 -ub 512 -t 12

Use the whole text file. Do not cut it to 179 physical lines: many rows contain multi-line documents, code, or math. With context 512, the saved Qwen3.5 run produced 179 calibration chunks.

Load it with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset(
    "text",
    data_files="bartowski_v5_semantic_25p.txt",
)
print(dataset["train"].num_rows)  # 519

License and attribution

This compact dataset keeps the original source attribution and MIT license information from lemon07r's Bartowski v5 semantic dataset. Please credit lemon07r when you redistribute or build on the source data.