datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
jepa-qwen3-32b-ce-cb-pra-2026-05-20
Qwen3-32B — CE-floor / Circuit-Breaker × Identity / MLP PRA
Trained LoRA adapters + attack results + benign evals for the Qwen3-32B PRA matrix (samuelsimko/AdversariaLLM experiments/configs/qwen3_32b_ce_cb.yaml @ commit 3a21c23), executed on a single 8×H200 tamia node.
Paper context: Predictive Representation Alignment Improves Generalization in LLM Safety (work in progress). PRA = a small predictor head that maps adversarial-view hidden states toward the clean view's commitment… See the full description on the dataset page: https://huggingface.co/datasets/memo-ozdincer/jepa-qwen3-32b-ce-cb-pra-2026-05-20.repro-causal-jepa-learning-world-models-through-object-level-latent-masking-traces
Agent traces
Agent sessions published from a Trackio Logbook.
jepadata
JepaData: 269K Adversarial Trajectories for LLM Safety
269,453 verified adversarial prompt–response trajectories for training and
evaluating jailbreak-resistant language models. JepaData spans eight attack
families and includes both harmful behaviors and benign prompts selected to
measure over-refusal.
The dataset was built as part of the large-scale adversarial data pipeline for
JEPA-Align and includes per-example attack, model, source, and transferability
provenance.… See the full description on the dataset page: https://huggingface.co/datasets/memo-ozdincer/jepadata.auto-jepa-v1
auto-jepa-v1
OpenCode session traces from the auto-jepa JEPA forecasting autoresearch runs.
Scope: last 10 orchestrator sessions per model (jepa-gemma-4-31b, jepa-qwen3.6-27b), each with its sub-agent tree.
How to browse
Dataset tab: searchable index (sessions_index.jsonl) — one row per session with path to the trace file.
Files tab: open *.trace.json for full OpenCode export, or cycle_manifest.yaml / model_manifest.yaml for per-cycle indexes.
Layout… See the full description on the dataset page: https://huggingface.co/datasets/dmnsh/auto-jepa-v1.jepa-cot-driving-benchchan-jepa-daily