jspace
Datasets
All datasets matching “jspace”jspace-unlearning
J-Access (J-space occupancy) × Machine Unlearning — 代码、lens、结果归档
归档日期:2026-08-22。对应论文草稿 paper/main_aaai.tex,数字权威来源
paper/MAINLINE_EXPERIMENTS.md(每个数字标注了来源 json)。
一句话:用 Jacobian lens(anthropics/jacobian-lens)在中间层读出"模型是否仍在
准备说出被遗忘的答案"(J-Access / JOcc),在 TOFU forget10 + Llama-3.2-1B-Instruct
的 398 个 OpenUnlearning 公开 checkpoint 上做三项研究:
Study 1 审计(残留普遍存在)、Study 2 预测(攻前占用 → 攻后复活,模型级成立/样本级失败)、
Study 3 优化压力(直接压制占用 = Goodhart,复活反而升高)。
1. 仓库内容
目录
内容
大小
src/
全部… See the full description on the dataset page: https://huggingface.co/datasets/Ziruibest/jspace-unlearning.jspace-hallucination-campaign
jspace hallucination campaign traces (Gemma-4-12B)
24,540 graded Gemma-4-12B responses with internal workspace features, from a
pre-registered cross-domain hallucination detection campaign. Companion to the
jspace repo and the
jspace-lenses model repo
(fitted Jacobian lenses plus the frozen error classifiers trained on this data).
Each row is one prompt, Gemma's greedy answer, an error label, and 31
deployable features read from the model's residual stream during generation
via… See the full description on the dataset page: https://huggingface.co/datasets/solarkyle/jspace-hallucination-campaign.jspace-w3-mediaeh-jspace-family-atlas
jspace-family-atlas -- aggregate exhaust
Aggregate-only: every file committed under this experiment's analysis-committed/ tree (dose-response tables, direction fits, gate AUROCs, manifests, and any other analysis artifact), copied byte-for-byte. No source question text, aliases, or per-row generation text -- analysis-committed/ never carries those.
HF repo: professorsynapse/eh-jspace-family-atlas
Provenance
Experiment: experiments/jspace-family-atlas
Amendment:… See the full description on the dataset page: https://huggingface.co/datasets/professorsynapse/eh-jspace-family-atlas.slm-jspace-tutorial-artifactseh-jspace-fresh-pool-census-qwen3-4b
J-space Fresh-Pool Census Qwen3-4B
This dataset is a public-safe row index and behavior-flag census for the
Epistemic Humility Research J-space layer-site replication.
It intentionally contains no question text, gold aliases, prompt text, model
generation text, hidden states, or row-level intervention outputs. It contains
only row identifiers, source/provenance fields, gold labels, selected behavioral
roles, and text-free baseline generation flags.
HF repo:… See the full description on the dataset page: https://huggingface.co/datasets/professorsynapse/eh-jspace-fresh-pool-census-qwen3-4b.