datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Rosetta-Activations
Rosetta Activations
Updated: 2026-06-15 02:30 UTC
Contrastive activation extractions for 17 semantic concepts across 46 language models,
supporting cross-architecture mechanistic interpretability research.
Companion concept pair corpus: jamesrahenry/Rosetta_Concept_Pairs
Papers: forthcoming
Dataset Structure
Rosetta-Activations/
├── rcp_v1/ # Current extraction line — richest data (N≈2000)
│ └── {Model_Name}/
│ ├── calibration_{concept}.npy… See the full description on the dataset page: https://huggingface.co/datasets/james-ra-henry/Rosetta-Activations.deception-probes-activations
Deception Probes Activations
Pre-extracted residual-stream activations for training and evaluating deception
detection probes on LLMs. Each example contains per-token hidden states from a
specific transformer layer, saved in bfloat16 safetensors format.
License
This dataset contains activations derived from multiple sources with different licenses.
See the LICENSE file for full details.
Component
Source
License
Apollo Probe Pairs (statements)
Azaria & Mitchell… See the full description on the dataset page: https://huggingface.co/datasets/xycoord/deception-probes-activations.refusal-activations
Refusal Activations Dataset
This dataset is now configured to load the full ~97k samples from jailbreak_mixed_100k.csv.
lists-activationssycophancy-activationsmetaphors-activationsthinking-model-activationsauditbench-activations-jlens-NLA
AuditBench activations, J-lens readouts and NLA verbalizations
Every token of every AuditBench prompt and every model response, from
meta-llama/Llama-3.3-70B-Instruct (revision 6f6073b423013f6a7d4d9f39144961bfbfbc386b) with one LoRA adapter per cell.
Responses were regenerated greedily and run to the model's own stopping point rather
than truncated at a fixed length, and the activations, readouts and verbalizations
cover the prompt as well as the response.
84 cells across 14… See the full description on the dataset page: https://huggingface.co/datasets/PranavViswanath/auditbench-activations-jlens-NLA.deception-activationsauthority-activationssandbagging-activationsolmo-activationsscience-activationsaction-atlas-groot-activations1M_activations_pile_10k_GPT_Gemma_Qwensycophancy-activationsqwen3-8b-activations-l20-l36
Qwen3 8B Activations for Layers 20 and 36
This dataset contains assistant-token residual activations harvested from Qwen/Qwen3-8B over 980000 training conversations from lmsys/lmsys-chat-1m.
We only generated for Layer 20 and 36 because each one costs 2TB and we simply cannot afford to store more :)
You can use this dataset to train SAEs, linear probes, other mech interp models etc, for Qwen3 8B.
We picked Qwen3 8B because this is a small part of a larger experiment to use feature… See the full description on the dataset page: https://huggingface.co/datasets/sammyliu/qwen3-8b-activations-l20-l36.deception-activationsmodel-inference-activationsactivationsbench-af-activations
Bench-AF: Alignment Faking Detection Activations & Probes
Activation caches and cross-validated linear probes for detecting alignment faking
in LLMs. Part of the Bench-AF research project.
Models
Model
Base
Adapter
llama-3-70b
Meta-Llama-3-70B-Instruct
None
llama-3-70b-base
Meta-Llama-3-70B
None
hal9000
Meta-Llama-3-70B-Instruct
bench-af/hal9000-adapter
pacifist
Meta-Llama-3-70B-Instruct
bench-af/pacifist-adapter
Datasets
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/LakshC/bench-af-activations.activations-and-barcodes
sveneziale/activations-and-barcodes
Compute artifacts pushed by tda-for-llms's Hugging-Face-backed pipeline
(hf.enabled: true in experiment.yaml).
Layout
Two top-level folders:
activations/{model_slug}/{corpus}/{revision}/{act_name}/
Raw per-cloud activation matrices extracted from the model, one independent
copy per checkpoint (revision). Independent of topology.metric — the
same activations are reused across every metric or topology config that… See the full description on the dataset page: https://huggingface.co/datasets/sveneziale/activations-and-barcodes.sae-activations-llama-3.1-8b-layer19-lmsys-chat-1m
SAE Feature Activations — Llama 3.1 8B Instruct, Layer 19 (LMSYS-Chat-1M)
This dataset contains Sparse Autoencoder (SAE) feature activations extracted from layer 19 of Meta's Llama 3.1 8B Instruct on conversations from LMSYS-Chat-1M.
It also has natural language explainations of features generated by GPT OSS 120B. See subset 4 for details.
The SAE used is Goodfire/Llama-3.1-8B-Instruct-SAE-l19, which decomposes layer-19 residual stream activations into interpretable sparse features.… See the full description on the dataset page: https://huggingface.co/datasets/scaleinvariant/sae-activations-llama-3.1-8b-layer19-lmsys-chat-1m.activations_and_barcodes_3108
sveneziale/activations_and_barcodes_3108
Compute artifacts pushed by tda-for-llms's Hugging-Face-backed pipeline
(hf.enabled: true in experiment.yaml).
Layout
Two top-level folders:
activations/{model_slug}/{corpus}/{revision}/{act_name}/
Raw per-cloud activation matrices extracted from the model, one independent
copy per checkpoint (revision). Independent of topology.metric — the
same activations are reused across every metric or topology config… See the full description on the dataset page: https://huggingface.co/datasets/sveneziale/activations_and_barcodes_3108.truth-probe-activationsopenwebtext-gemma3-tokenized-1024-activations-layer23
OpenWebText — Gemma-3-1B Hidden State Activations (Layer 23)
Precomputed hidden state activations before layer 23 of Gemma-3-1B-IT for the OpenWebText dataset, tokenized with sequence length 1024.
Designed for training a Titans memory layer that replaces layer 23 of Gemma 3.
Dataset Structure
Each example contains the inputs to layer 23:
Field
Shape
Dtype
Description
activations
(1024, 1152)
float32
Hidden state activations (cast from bfloat16)
mask(1024… See the full description on the dataset page: https://huggingface.co/datasets/veriga/openwebtext-gemma3-tokenized-1024-activations-layer23.nla-activations-llama-70b-layer53
NLA Activations — Llama-3.3-70B Layer 53
Layer-53 residual-stream activations extracted from all 84 model organisms in the auditing-agents audit benchmark. These are the raw floating-point vectors fed into the NLA (Natural Language Autoencoder) actor to produce natural-language descriptions of what each model represents at each token position.
The companion verbalization dataset (activations → text) is at Realmbird/nla-av-responses-llama-70b-layer53.
What's in this… See the full description on the dataset page: https://huggingface.co/datasets/Realmbird/nla-activations-llama-70b-layer53.diffusion_policy_robocasa_activations_latest_chkpt
Diffusion Policy — RoboCasa Activations (latest checkpoint)
Per-step, per-episode activation traces collected from a DiffusionTransformerHybridImagePolicy (the diffusion_policy library) rolled out on RoboCasa benchmark tasks. Captured with collect_activations_robocasa.py at the latest training checkpoint.
These traces are the input expected by the conceptor / SAE steering pipelines under diffusion_policy/experiments/robocasa_steering/ and diffusion_policy/experiments/sae/ — see… See the full description on the dataset page: https://huggingface.co/datasets/brandonyang/diffusion_policy_robocasa_activations_latest_chkpt.latenet-v0-activations-llama3.1-70b-base
meta-llama/Llama-3.1-70B — Activation Dataset
Cached activations extracted from meta-llama/Llama-3.1-70B (revision 349b2ddb53ce8f2849a6c168a81980ab25258dac).
Full-sequence activations (80 layers, 8192 dim, float16, all tokens) from meta-llama/Llama-3.1-70B (base) on 23724 LateNet v0 statements (affirmative + negated). Extracted via NDIF. Raw statements only (no chat template). Prompts ordered by negated→generator→pair_id for contiguous domain shards.
Contents… See the full description on the dataset page: https://huggingface.co/datasets/alliedtoasters/latenet-v0-activations-llama3.1-70b-base.action-atlas-smolvla-activations