Exulan/qwen3-8b-pol-lora
Text Generation • Updated • 3
Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Contains pervasive slurs, hate speech, extremist rhetoric, graphic language,
harassment, misinformation, and potentially residual identifying text.
Read TERMS.md before requesting access.
Log in or Sign Up to review the conditions and access this dataset content.
Packed /pol/ threads and greentext
Training-ready token blocks derived from the 4plebs /pol/ archive covering
2014 through 2025. This is the exact 40% greentext / 60% reconstructed-thread
mixture used for the Qwen3-8B LoRA run; it is not the complete raw archive.
Companion model
The finished companion adapter is available at Exulan/qwen3-8b-pol-lora. It contains the verified final 3M-tier guided LoRA selected after matched comparison against the 1M tier. This repository contains the raw packed phase shown below; the later 3M guidance cohort and replay mixture are not distributed as part of this dataset release.
Content warning
The data was not filtered for ideology, profanity, toxicity, or slurs. It can contain hate speech, extremist material, graphic content, harassment, misinformation, and identifying strings missed by automated redaction. It is not suitable for minors or an unmoderated public service.
Data
| Split | Blocks | Valid tokens | Supervised tokens |
|---|---|---|---|
| train | 397,605 | 356,198,483 | 347,150,490 |
| validation | 4,160 | 3,695,928 | 3,601,536 |
| test | 1,897 | 1,680,898 | 1,639,495 |
Every block is 1,024 Qwen3 tokens. Each split contains:
input_ids.bin: little-endianuint32labels.bin: little-endianint32; ignored positions are-100lengths.bin: one little-endianuint32valid length per block
data/manifest.json records the tokenizer revision, counts, mixture, and
windowing policy. packed_dataset.py provides a memory-mapped PyTorch loader.
from packed_dataset import PackedMemmapDataset
train = PackedMemmapDataset(
"data/train", expected_sequence_length=1024
)
print(len(train), train[0].keys())
Install the loader dependencies with pip install -r requirements.txt.
Small decoded examples are in samples/; the binary payload is canonical.
Source and rights
Source archive: https://archive.4plebs.org/pol/
Archived posts are third-party material and are not relicensed here. Original release metadata and documentation are dedicated under CC0 where possible. See TERMS.md.
- Downloads last month
- 18
Models trained or fine-tuned on Exulan/pol-2014-2025-packed