Datasets:

Beicicc
/

probeshift-activation-cache

model string	dataset string	distribution string	pooling string	n int64	n_layers int64	hidden int64	dtype string
gpt2-medium	ag_news	train	mean	1,500	25	1,024	float16
gpt2-medium	amazon_cf	train	mean	1,500	25	1,024	float16
gpt2-medium	cola	train	mean	1,500	25	1,024	float16
gpt2-medium	counterfact	train	mean	1,500	25	1,024	float16
gpt2-medium	dbpedia	train	mean	1,500	25	1,024	float16
gpt2-medium	emotion	train	mean	1,500	25	1,024	float16
gpt2-medium	imdb	train	mean	1,500	25	1,024	float16
gpt2-medium	spam	train	mean	1,500	25	1,024	float16
gpt2-medium	sst2	train	mean	1,500	25	1,024	float16
gpt2-medium	stance	train	mean	389	25	1,024	float16
gpt2-medium	subj	train	mean	1,500	25	1,024	float16
gpt2-medium	tweet_hate	train	mean	1,500	25	1,024	float16
gpt2-medium	tweet_irony	train	mean	1,500	25	1,024	float16
gpt2-medium	tweet_offensive	train	mean	1,500	25	1,024	float16
gpt2	ag_news	train	mean	1,500	13	768	float16
gpt2	amazon_cf	train	mean	1,500	13	768	float16
gpt2	cola	train	mean	1,500	13	768	float16
gpt2	counterfact	train	mean	1,500	13	768	float16
gpt2	dbpedia	train	mean	1,500	13	768	float16
gpt2	emotion	train	mean	1,500	13	768	float16
gpt2	imdb	train	mean	1,500	13	768	float16
gpt2	spam	train	mean	1,500	13	768	float16
gpt2	sst2	train	mean	1,500	13	768	float16
gpt2	stance	train	mean	389	13	768	float16
gpt2	subj	train	mean	1,500	13	768	float16
gpt2	tweet_hate	train	mean	1,500	13	768	float16
gpt2	tweet_irony	train	mean	1,500	13	768	float16
gpt2	tweet_offensive	train	mean	1,500	13	768	float16
pythia-1.4b	ag_news	train	mean	1,500	25	2,048	float16
pythia-1.4b	amazon_cf	train	mean	1,500	25	2,048	float16
pythia-1.4b	cola	train	mean	1,500	25	2,048	float16
pythia-1.4b	counterfact	train	mean	1,500	25	2,048	float16
pythia-1.4b	dbpedia	train	mean	1,500	25	2,048	float16
pythia-1.4b	emotion	train	mean	1,500	25	2,048	float16
pythia-1.4b	imdb	train	mean	1,500	25	2,048	float16
pythia-1.4b	spam	train	mean	1,500	25	2,048	float16
pythia-1.4b	sst2	train	mean	1,500	25	2,048	float16
pythia-1.4b	stance	train	mean	389	25	2,048	float16
pythia-1.4b	subj	train	mean	1,500	25	2,048	float16
pythia-1.4b	tweet_hate	train	mean	1,500	25	2,048	float16
pythia-1.4b	tweet_irony	train	mean	1,500	25	2,048	float16
pythia-1.4b	tweet_offensive	train	mean	1,500	25	2,048	float16
pythia-160m	ag_news	train	mean	1,500	13	768	float16
pythia-160m	amazon_cf	train	mean	1,500	13	768	float16
pythia-160m	cola	train	mean	1,500	13	768	float16
pythia-160m	counterfact	train	mean	1,500	13	768	float16
pythia-160m	dbpedia	train	mean	1,500	13	768	float16
pythia-160m	emotion	train	mean	1,500	13	768	float16
pythia-160m	imdb	train	mean	1,500	13	768	float16
pythia-160m	spam	train	mean	1,500	13	768	float16
pythia-160m	sst2	train	mean	1,500	13	768	float16
pythia-160m	stance	train	mean	389	13	768	float16
pythia-160m	subj	train	mean	1,500	13	768	float16
pythia-160m	tweet_hate	train	mean	1,500	13	768	float16
pythia-160m	tweet_irony	train	mean	1,500	13	768	float16
pythia-160m	tweet_offensive	train	mean	1,500	13	768	float16
pythia-410m	ag_news	train	mean	1,500	25	1,024	float16
pythia-410m	amazon_cf	train	mean	1,500	25	1,024	float16
pythia-410m	cola	train	mean	1,500	25	1,024	float16
pythia-410m	counterfact	train	mean	1,500	25	1,024	float16
pythia-410m	dbpedia	train	mean	1,500	25	1,024	float16
pythia-410m	emotion	train	mean	1,500	25	1,024	float16
pythia-410m	imdb	train	mean	1,500	25	1,024	float16
pythia-410m	spam	train	mean	1,500	25	1,024	float16
pythia-410m	sst2	train	mean	1,500	25	1,024	float16
pythia-410m	stance	train	mean	389	25	1,024	float16
pythia-410m	subj	train	mean	1,500	25	1,024	float16
pythia-410m	tweet_hate	train	mean	1,500	25	1,024	float16
pythia-410m	tweet_irony	train	mean	1,500	25	1,024	float16
pythia-410m	tweet_offensive	train	mean	1,500	25	1,024	float16
pythia-6.9b	ag_news	train	mean	1,500	33	4,096	float16
pythia-6.9b	counterfact	train	mean	1,500	33	4,096	float16
pythia-6.9b	dbpedia	train	mean	1,500	33	4,096	float16
pythia-6.9b	emotion	train	mean	1,500	33	4,096	float16
pythia-6.9b	sst2	train	mean	1,500	33	4,096	float16
pythia-6.9b	subj	train	mean	1,500	33	4,096	float16
pythia-6.9b	tweet_hate	train	mean	1,500	33	4,096	float16
pythia-70m	ag_news	train	mean	1,500	7	512	float16
pythia-70m	amazon_cf	train	mean	1,500	7	512	float16
pythia-70m	cola	train	mean	1,500	7	512	float16
pythia-70m	counterfact	train	mean	1,500	7	512	float16
pythia-70m	dbpedia	train	mean	1,500	7	512	float16
pythia-70m	emotion	train	mean	1,500	7	512	float16
pythia-70m	imdb	train	mean	1,500	7	512	float16
pythia-70m	spam	train	mean	1,500	7	512	float16
pythia-70m	sst2	train	mean	1,500	7	512	float16
pythia-70m	stance	train	mean	389	7	512	float16
pythia-70m	subj	train	mean	1,500	7	512	float16
pythia-70m	tweet_hate	train	mean	1,500	7	512	float16
pythia-70m	tweet_irony	train	mean	1,500	7	512	float16
pythia-70m	tweet_offensive	train	mean	1,500	7	512	float16
qwen2.5-0.5b	ag_news	train	mean	1,500	25	896	float16
qwen2.5-0.5b	amazon_cf	train	mean	1,500	25	896	float16
qwen2.5-0.5b	cola	train	mean	1,500	25	896	float16
qwen2.5-0.5b	counterfact	train	mean	1,500	25	896	float16
qwen2.5-0.5b	dbpedia	train	mean	1,500	25	896	float16
qwen2.5-0.5b	emotion	train	mean	1,500	25	896	float16
qwen2.5-0.5b	imdb	train	mean	1,500	25	896	float16
qwen2.5-0.5b	spam	train	mean	1,500	25	896	float16
qwen2.5-0.5b	sst2	train	mean	1,500	25	896	float16

End of preview.

ProbeShift Activation Cache

Residual-stream activations backing the ProbeShift benchmark — a label-free study of linear-probe direction stability under label-preserving semantic shift. Ships so the benchmark's numbers reproduce in minutes (no re-extraction needed).

Layout

cache_seed{0..4}/<model>/<dataset>/<distribution>/
    acts.npy     float16  [N, L+1, H]   masked-mean-pooled residual stream (L+1 = embeddings + L layers)
    labels.npy   int64    [N]
    ids.npy      int64    [N]            stable example ids (align across distributions)
    meta.json    {model, dataset, distribution, pooling, n, n_layers, hidden}

models (8): pythia-70m/160m/410m/1.4b/6.9b, gpt2, gpt2-medium, qwen2.5-0.5b
datasets (14 → 12 concepts): sst2, imdb (sentiment); ag_news, dbpedia (topic); counterfact (truth); emotion; tweet_hate; tweet_irony; tweet_offensive; subj (subjectivity); spam; cola (grammaticality); stance; amazon_cf (counterfactual)
distributions: train, iid, paraphrase/domain/length (label-preserving OOD shifts), aug0/aug1/aug2 (de/fr/ru back-translation augmentations)
seeds: 0–4 (each an independent example draw + independent paraphrase — Option A replication)

Load

import numpy as np
acts = np.load("cache_seed0/pythia-410m/sst2/iid/acts.npy", mmap_mode="r")  # [N, L+1, H] fp16
labels = np.load("cache_seed0/pythia-410m/sst2/iid/labels.npy")

Valid configurations

The paper's paraphrase × LogReg grid has 472 valid configurations: 7 models (pythia-70m/160m/410m/1.4b, gpt2, gpt2-medium, qwen2.5-0.5b) × 14 datasets × 5 seeds = 490, minus 18 cells skipped for massive activations. The pythia-6.9b shards (seed 0) back a preliminary spot-check. The 472 valid keys are those in the supplementary package's records/predictors.jsonl.

Paper and code

Kun Zhang and Chunwei Xia. ProbeShift: Predicting Probe-Direction Rotation from Stability is Largely a Sampling-Noise Floor. Proceedings of the 18th Asian Conference on Machine Learning (ACML 2026), PMLR. The per-configuration records and all analysis scripts ship as the paper's supplementary material (supplement/README.md). Activations were produced on RTX 4090 GPUs (≤200 GPU·h, $0 API, zero new annotation).

Downloads last month: 2,309

Total file size:

158 GB