keptn/quant-calibration-v1

Quant calibration v1

A fixed, public calibration set for weight quantization (GPTQ, AWQ, imatrix). It has 1,213 conversations with about 1.5 million tokens. It contains chat in 10 languages, tasks on long documents, code, math with reasoning traces and tool sessions. Each row is a message list, so each model renders it with its own chat template.

Rows	1,213
Tokens	1,516,060 (Gemma 4 tokenizer, message text without template)
Languages	English 57 %, German 30 %, 8 other European languages 13 % (by tokens)
Rows with a reasoning trace	154
Rows with tool definitions	121
Version	1.0 (2026-10-30)

Purpose

Quantization measures which weights matter on real text. The calibration text must look like the traffic that the model will see. This set is made for chat and agent models. It is not an evaluation set. Do not use it to measure quality.

The set is fixed. Every model gets the same rows. This makes calibrations repeatable and comparable.

Contents

subset	rows	tokens	source	license
chat-oasst2	497	180,023	OpenAssistant/oasst2, best-ranked path per tree; human answers	Apache-2.0
chat-aya	323	90,156	CohereLabs/aya_dataset, 1 to 3 pairs of one language per row; human answers	Apache-2.0
documents	99	365,993	task prompts (summarize, extract, explain, translate) on excerpts of 34 public documents: Wikipedia, public-domain books, German laws and parliament transcripts, EUR-Lex, US federal works	per row, see below
code-files	95	196,207	task prompts (review, explain, test, refactor) on files of radiance v1.12.0 and the CPython 3.12 standard library	Apache-2.0, PSF-2.0
math	27	106,659	nvidia/Nemotron-Post-Training-Dataset-v1, math split, with reasoning trace	CC-BY-4.0
stem	51	123,371	Nemotron-Post-Training-Dataset-v1, stem split, with reasoning trace	CC-BY-4.0
tools-nemotron	76	255,480	Nemotron-Post-Training-Dataset-v1, tool_calling split: tool definitions, calls, results, answers	CC-BY-4.0
tools-agent	45	198,171	synthetic agent sessions made for this set: read, search and shell calls; the tool results are real file contents and git logs of the code sources above	CC-BY-4.0; tool results under their source license

The documents and code-files rows end with a user turn. They have no answer, because calibration only needs the prefill of these prompts. The chat, math, stem and tools rows contain full answers.

Tokens by language: English 867,047 · German 456,094 · French 34,660 · Portuguese 36,782 · Spanish 27,129 · Italian 25,167 · Polish 22,658 · Catalan 21,131 · Dutch 14,195 · Swedish 11,197.

What is not in the set

• No private chat logs and no data from users of chat services. For this reason the Nemotron v1 chat split (prompts from lmsys-chat-1m) is not used.
• No Nemotron v1 code rows. In that split the task text is only a placeholder ("-").
• No test or validation splits of common benchmarks.

Contamination check

A row is removed if it shares a sequence of 13 words with one of these sets:

set	use
MT-Bench questions and reference answers	KL measurement of quantized models
GSM8K test	benchmark
IFEval	benchmark
wikitext-2 test	perplexity

The check removed 0 rows. MT-Bench turns with fewer than 13 words were also searched as exact text: 0 hits. Other benchmarks (for example MATH, GPQA or LiveCodeBench) were not checked.

Format

One JSON object per line in data/train.jsonl:

field	content
id	row id
category, subset	chat, documents, code, math, stem or tools; subset as in the table above
language	ISO code
thinking	true if an assistant turn has a reasoning trace
tokens_gemma4	token count of the message text (Gemma 4 tokenizer, no template)
source, license	source and license of this row
tools	tool definitions as a JSON string (OpenAI format), or null
messages	role, content; optional reasoning_content, tool_calls (arguments as a JSON string), name

sources.json lists every source with its revision. For each document it gives the URL, the Wikipedia revision id where it applies, the license and the SHA-256 of the full text.

Use

render.py renders the set with the chat template of a model:

pip install transformers jinja2
python render.py --model google/gemma-4-31B-it --data data/train.jsonl --out calib-gemma4.jsonl

• Rows that end with a user or tool turn get the generation prompt.
• Rows with a reasoning trace render with enable_thinking=True. Of the other rows, 30 % render with thinking on (--thinking-share).
• Rows that the template does not accept are skipped and counted.

Tested with the templates of Gemma 4 31B, Nemotron 3.5 Bolt 74B and Ling 3.0 flash: all 1,213 rows render.

For llama.cpp llama-imatrix, join the prompt fields of the output into one text file.

Limits

• 1.5 million tokens is enough for GPTQ Hessians of dense models and of most mixture-of-experts models. Rare experts can get few tokens.
• The math, stem and tool rows are English only.
• There is no image, audio or video data.
• The answers in the Nemotron rows come from Qwen3-235B-A22B and DeepSeek-R1-0528. NVIDIA notes that models trained on these outputs can be subject to the Qwen and DeepSeek licenses.

Licenses

Each row has its own license field. Each part keeps its own license. In the tools-agent rows the session is CC-BY-4.0 and each tool result keeps the license of its source (also in the license field).

license	rows
Apache-2.0	873
CC-BY-4.0	199
public domain (old books, German official works, US federal works)	73
PSF-2.0	42
CC-BY-SA-4.0 (Wikipedia; share alike)	19
EUR-Lex reuse (Decision 2011/833/EU)	7

Credits

• OpenAssistant (oasst2), Cohere Labs (Aya dataset), NVIDIA (Nemotron-Post-Training-Dataset-v1).
• Wikipedia authors, Deutscher Bundestag, Bundesverfassungsgericht, gesetze-im-internet.de, EUR-Lex, and the US federal agencies whose works are used.
• radiance: StillDeadcode. CPython: Python Software Foundation.