reneeice/data-ood-l-hardneg

OOD Hard Negative Mining Dataset

Dataset for fine-tuning OOD (out-of-distribution) detection on AI-generated text.

Files

File	Size	Description
cc_hard_negatives.pickle	11 MB	12,000 new CC (CommonCrawl) human text samples from 4 different CC-MAIN dumps. Quality-filtered for natural prose.
cc_hard_neg_scores.npz	193 KB	Per-text scores from the current OOD model on the 12K CC samples. Used to identify hard negatives (score > 0.3).
cc_training_data.pickle	5 MB	Original 5,000 CC human samples (from CC-MAIN-2023-50).
mixed_ood_cc_v2.pickle	327 MB	Combined training dataset: 79,759 samples (45,878 human + 33,881 AI). Includes original Pile + original CC + new CC with hard negative duplication.

Data Format

Each pickle contains a list of tuples: (text: str, score: float, label: int)

• label=0 = human/ID text
• label=1 = AI/OOD text
• score is unused (always 0.0)

Hard Negative Details

The 12K CC samples were scored with the current OOD model. 541/12000 (4.5%) scored > 0.3 despite being human text. These hard negatives were duplicated 3x in the training mix to emphasize them.

Failure Mode Categories:

1. Code/technical snippets (~1%): Lisp, Python, CSS code that looks structured like AI
2. Blog nav/UI text (~1%): Archive pages with date trees, link lists
3. Marketing/spam (~1%): Game cheats, product descriptions (may actually be AI)
4. Template text (~1%): Wedding officiant, boilerplate services
5. Genuine natural text (~0.5%): Gardening blogs, author bios — true false positives

Usage

import pickle
data = pickle.load(open('mixed_ood_cc_v2.pickle', 'rb'))
texts = [d[0] for d in data]
labels = [d[2] for d in data]