Datasetpublic

DarrenJiaImbue/ai-detection-demo-gemma-logits

ai-detection-demo — Gemma 4 E4B logits Pre-computed last-token vocabulary logits from int4-quantized google/gemma-4-E4B-it, applied over DarrenJiaImbue/ai-detection-demo-dataset. These logits are the training input to DarrenJiaImbue/ai-detection-demo-gemma_4_e4b, the linear-probe classification head. Structure ├── train/ │ ├── logits.npy (14000, 262144) fp16 │ ├── labels.npy (14000,) int64 — cosine-distance bucket 0..3 │ └── meta.json ├── val/ │… See the full description on the dataset page: https://huggingface.co/datasets/DarrenJiaImbue/ai-detection-demo-gemma-logits.

sourceHugging Facecc-by-nc-sa-4.0updated 3mo agoView on Hugging Face

0likes41downloads

Dataset Card

ai-detection-demo — Gemma 4 E4B logits

Pre-computed last-token vocabulary logits from int4-quantized `google/gemma-4-E4B-it`, applied over `DarrenJiaImbue/ai-detection-demo-dataset`.

These logits are the training input to `DarrenJiaImbue/ai-detection-demo-gemma_4_e4b`, the linear-probe classification head.

Structure

├── train/
│   ├── logits.npy      (14000, 262144) fp16
│   ├── labels.npy      (14000,) int64  — cosine-distance bucket 0..3
│   └── meta.json
├── val/
│   ├── logits.npy      (1742, 262144) fp16
│   ├── labels.npy      (1742,) int64
│   ├── meta.json
│   └── sample.parquet  — text_id + text + source columns for row alignment
└── test/
    ├── logits.npy      (2500, 262144) fp16
    ├── labels.npy      (2500,) int64
    ├── meta.json
    └── sample.parquet

Class balance: train is perfectly balanced (3,500 per bucket). Val has 500/500/242/500. Test is balanced (625 per bucket).

Split rows are a subset of the corresponding split in ai-detection-demo-dataset. The sample.parquet for val and test carries the original text_id and can be joined back to the parent dataset.

Loading

python

import numpy as np
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="DarrenJiaImbue/ai-detection-demo-gemma-logits",
    repo_type="dataset",
)
train_logits = np.load(f"{root}/train/logits.npy", mmap_mode="r")   # (14000, 262144) fp16
train_labels = np.load(f"{root}/train/labels.npy")                  # (14000,) int64

Label convention

labels.npy is an integer bucketing of each row's cosine distance (supervision signal from the parent dataset):

bucket	cosine range
0	`score ≤ 0.03`
1	`0.03 < score < 0.09`
2	`0.09 ≤ score < 0.15`
3	`score ≥ 0.15`

License

CC BY-NC-SA 4.0. Non-commercial research use only. Derivative redistributions must use the same license.