---
license: other
license_name: various-open
license_link: https://huggingface.co/datasets/PleIAs/common_corpus
task_categories:
  - text-generation
tags:
  - quantization
  - awq
  - calibration
  - moe
  - qwen3.5
size_categories:
  - n<1K
source_datasets:
  - PleIAs/common_corpus
---

# Qwen3.5 MoE AWQ Calibration Dataset

Calibration dataset for AWQ (Activation-Aware Weight Quantization) of
[Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) and
[Qwen/Qwen3.5-35B-A3B-Base](https://huggingface.co/Qwen/Qwen3.5-35B-A3B-Base).

Designed for MoE expert routing diversity: Qwen3.5-35B-A3B has 256 experts with 8
active per token, so calibration data needs broad domain coverage to exercise as many
routing paths as possible.

## Sampling methodology

**Source**: [PleIAs/common_corpus](https://huggingface.co/datasets/PleIAs/common_corpus)
(open multi-domain corpus with labeled collections)

**Filtering**:
- Token count between 512 and 8192 (long enough to be representative, not wastefully long)
- Excluded 12 collections with known quality issues:
  - OCR artifacts: Anno, BNL Newspapers (1841-1879), Deutsches Zeitungsportal,
    German-PD-Newspapers, Spanish-PD-Newspapers, French-PD-Newspapers, Italian-PD,
    Marianne-Europe, NewZealand-PD-Newspapers
  - Structured metadata (not natural text): Wikidata, Eurovoc
  - Encoding artifacts: French-Science-Pile

**Stratified sampling** (256 samples, seed=42):
- **25% code** (64 samples): Github Open Source + StackExchange, stratified by
  programming language (64 languages represented)
- **75% text** (192 samples): Round-robin across 83 non-code
  collections (government, legal, scientific, books, multilingual, etc.)
- Bucket keys are sorted then shuffled (deterministic with seed) so no source is
  systematically excluded by alphabetical ordering

**Result**: 85 collections, 64 code languages, 22 natural languages

## Statistics

**Token counts**: min=515, median=1352, max=8143, mean=2278

### Collections

| Collection | Samples |
|---|---|
| Github Open Source | 48 |
| StackExchange | 16 |
| Portuguese-PD | 3 |
| Caselaw Access Project | 3 |
| French-PD-diverse | 3 |
| GATT_library | 3 |
| Wiki Discussions | 3 |
| Câmara dos Deputados – Discursos | 3 |
| Spanish-Science-Pile | 3 |
| OpenAlex | 3 |
| Japanese-PD | 3 |
| German-PD | 3 |
| domsdatabasen | 3 |
| OpenITI | 3 |
| health_hovedstaden | 3 |
| Polish-PD | 3 |
| RusLawOD | 3 |
| Spanish-PD-Books | 3 |
| Eurlex | 3 |
| TEDEUTenders | 3 |
| US-PD-Books | 3 |
| Open-Science-Pile | 3 |
| ecfr | 3 |
| Open Legal Data | 3 |
| Pralekha | 3 |
| Superior Tribunal de Justiça | 3 |
| Entscheidungen des Bundesverwaltungsgerichts | 3 |
| enevaeldens_nyheder | 3 |
| LoC-PD-Books | 3 |
| Entscheidungen des Bundesgerichtshofs | 3 |
| Youtube-Commons | 3 |
| Creative Commons Common Crawl (CCC) | 3 |
| Wikipedia | 3 |
| usc | 3 |
| UK Hansard – Public Bill Committees | 3 |
| Open Korean Historical Corpus - The Records of Daily Reflections (Ilseongnok) | 3 |
| Open Korean Historical Corpus - Korean Modern and Contemporary Magazines (한국근현대잡지자료) | 2 |
| Entscheidungen des Bundesfinanzhofs | 2 |
| Canadian Hansard | 2 |
| Amtliche Entscheidungssammlung des Bundesverfassungsgerichts | 2 |
| UK Hansard – Westminster Hall | 2 |
| Court Listener | 2 |
| UK Hansard – Scottish Parliament | 2 |
| Open Korean Historical Corpus - Diaries of the Royal Secretariat (Seungjeongwon ilgi) | 2 |
| French-PD-Books | 2 |
| Entscheidungen des Bundesgerichtshofs in Strafsachen aus dem 20. Jahrhundert | 2 |
| English-PD | 2 |
| Open Australian Legal Corpus | 2 |
| Open Korean Historical Corpus - Naver News Library | 2 |
| Wikisource | 2 |
| USPTO | 2 |
| Europeana | 2 |
| WTO | 2 |
| reg_docs | 2 |
| UK Hansard – House of Commons | 2 |
| German-Science-Pile | 2 |
| SEC | 2 |
| Latin-PD | 2 |
| Chinese-Court-Decisions | 2 |
| arXiv | 2 |
| UK Hansard – Written Answers | 2 |
| fr | 2 |
| UK Hansard – Lords Written Statements | 2 |
| SBB Fulltexts | 2 |
| UK Hansard – Northern Ireland Assembly | 2 |
| dockets | 2 |
| Entscheidungen des Bundespatentgerichts | 2 |
| Korean National Law Information Center | 2 |
| UK Hansard – House of Lords | 2 |
| fdlp | 2 |
| Open Korean Historical Corpus - Korean Literary Collections | 2 |
| UN-Digital-Library | 2 |
| govinfo | 2 |
| dotgov | 2 |
| French Open Data | 2 |
| ai-aktindsigt | 2 |
| US-PD-Newspapers | 2 |
| Gutenberg | 1 |
| Open Korean Historical Corpus - Veritable Records of the Joseon Dynasty | 1 |
| Open Korean Historical Corpus - Korean Newspaper Archive | 1 |
| UK Hansard – Lords Written Answers | 1 |
| Open Korean Historical Corpus - Gaksadeungnok | 1 |
| ukleg | 1 |
| Czech-PD | 1 |
| Multilingual-PD | 1 |

### Code languages

| Language | Samples |
|---|---|
| Dutch | 1 |
| Inno Setup | 1 |
| Gradle | 1 |
| ECL | 1 |
| Polish | 1 |
| Makefile | 1 |
| Rust | 1 |
| French | 1 |
| Lithuanian | 1 |
| Objective-C | 1 |
| Spanish | 1 |
| GDScript | 1 |
| Squirrel | 1 |
| Dockerfile | 1 |
| Shell | 1 |
| Swahil | 1 |
| Ada | 1 |
| PHP | 1 |
| Hack | 1 |
| Mathematica | 1 |
| Turtle | 1 |
| Somali | 1 |
| YANG | 1 |
| AsciiDoc | 1 |
| EJS | 1 |
| Hausa | 1 |
| Unity3D Asset | 1 |
| Portuguese | 1 |
| Croatian | 1 |
| Less | 1 |
| Metal | 1 |
| Nim | 1 |
| Assembly | 1 |
| Swift | 1 |
| Julia | 1 |
| Gerber Image | 1 |
| German | 1 |
| Crystal | 1 |
| Sardinian | 1 |
| JavaScript | 1 |
| HTML+Razor | 1 |
| G-code | 1 |
| Smali | 1 |
| OpenType Feature File | 1 |
| MATLAB | 1 |
| Sass | 1 |
| Vue | 1 |
| Lua | 1 |
| Dart | 1 |
| SCSS | 1 |
| Pascal | 1 |
| Jupyter Notebook | 1 |
| FreeBasic | 1 |
| Norwegian Nynorsk | 1 |
| Vim Script | 1 |
| Swedish | 1 |
| PLpgSQL | 1 |
| RDoc | 1 |
| Norwegian Bokmål | 1 |
| Go | 1 |
| Lean | 1 |
| TSX | 1 |
| Friulian | 1 |
| TSV | 1 |

### Natural languages (top 20)

| Language | Samples |
|---|---|
| English | 86 |
| German | 31 |
| French | 12 |
| Danish | 11 |
| Hanmun | 8 |
| Portuguese | 6 |
| Modern Korean | 5 |
| Spanish | 5 |
| Russian | 4 |
| Japanese | 3 |
| Arabic | 3 |
| Polish | 3 |
| Portugueuse | 2 |
| Urdu | 2 |
| Chinese | 2 |
| Czech | 2 |
| Korean | 2 |
| Latvian | 1 |
| Unknown | 1 |
| Thai | 1 |
| *(+2 more)* | |

## Usage

```python
from datasets import load_from_disk

ds = load_from_disk("qwen3.5-moe-awq-calibration/data")
```

Or with the quantization script:

```python
from datasets import load_dataset

ds = load_dataset("<repo_id>", split="train")
```

## License

Sampled from [PleIAs/common_corpus](https://huggingface.co/datasets/PleIAs/common_corpus),
which aggregates openly licensed text under various open licenses (635 distinct licenses
across the full corpus). Each record retains its original license. See the `license` field
on individual records and the source dataset for details.