datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
qwen3-8b-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.qwen2.5-7b-instruct-nla-L20-finefineweb-100k
Qwen2.5-7B-Instruct NLA training data — residual stream, block 20
Training data for a Natural Language Autoencoder on Qwen/Qwen2.5-7B-Instruct:
residual-stream activations paired with natural-language explanations of the text
they were taken from.
Unlike the dataset this is derived from, the activation_vector column is
included — every EasyNLA/nanoNLA trainer requires it.
Trained models: https://huggingface.co/Yooniel/qwen2.5-7b-instruct-nla-L20
(AV val ppl 4.07, AR held-out FVE… See the full description on the dataset page: https://huggingface.co/datasets/Yooniel/qwen2.5-7b-instruct-nla-L20-finefineweb-100k.L20diffing-stats-gemma-2-9b-it-L20-k100-lr1e-04-Crosscoderdiffing-stats-gemma-2-9b-it-DPO-L20-k100-lr1e-04-dpo-simpo-Crosscoderdiffing-stats-gemma-2-9b-L20-k100-lr1e-04-base-it-Crosscodernla-qwen2.5-7b-L20-results-famous-ads
Famous Ad Slogans — NLA Activation Analysis (Qwen2.5-7B, Layer 20)
What does a language model internally represent when it reads a famous
advertising slogan? This dataset answers that for ~150 well-known slogans by
decoding the model's residual-stream activation into plain English with a
Natural Language Autoencoder (NLA).
slogan ──(Qwen2.5-7B-Instruct)──► layer-20 activation (final token)
──(AV: Activation Verbalizer)──► English description of the vector
──(AR:… See the full description on the dataset page: https://huggingface.co/datasets/SolshineMisfit/nla-qwen2.5-7b-L20-results-famous-ads.diffing-stats-gemma-2-9b-it-L20-mu1.0e-01-lr1e-04-local-shuffling-CrosscoderLossdiffing-stats-gemma-2-9b-L20-k100-lr1e-04-Crosscodergemma4-26b-a4b-nla-L20-finefineweb-100k-fp32
gemma-4-26B-A4B NLA warmstart data (L20, fp32 golds, FineFineWeb 100k)
Gemma-4 analog of ceselder/qwen3-8b-nla-L24-finefineweb-100k: same prefixes and
gold Claude explanations, activations re-harvested from gemma-4-26B-A4B-it layer 20
in fp32 (bf16 MoE activations are kernel-shape-dependent; fp32 is exact).
Sidecar .nla_meta.yaml files included per parquet.
details_2A2I-R__L200MT
Dataset Card for Evaluation run of 2A2I-R/L200MT
Dataset automatically created during the evaluation run of model 2A2I-R/L200MT.
The dataset is composed of 136 configuration, each one coresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional configuration… See the full description on the dataset page: https://huggingface.co/datasets/OALL/details_2A2I-R__L200MT.diffing-stats-gemma-2-9b-L20-k100-lr1e-04-base-dpo-Crosscoderdiffing-stats-gemma-2-9b-it-L20-k100-lr1e-04-CCLossdiffing-stats-gemma-2-9b-it-SimPO-L20-mu1.0e-01-lr1e-04-CrosscoderLossdiffing-stats-gemma-2-9b-it-L20-mu4.0e-02-lr1e-04-CrosscoderLossnlt-l16-l20-archive
Natural-Language Transcoder — Results (2026-07-28/29)
RL-only conversion of the released Qwen2.5-7B NLA pair (kitft/nla-qwen2.5-7b-L20)
into a layer-16 → layer-20 transcoder. 300 GRPO steps, absolute mode, warm-started
from the released AV (DCP-converted) + released AR. Code: public repo
Anwen0507/natural_language_transcoders, branch transcoder.
Headline table — held-out eval, 512 rows, greedy, direction-MSE = 2(1−cos)
System
MSE
cos
Identity baseline… See the full description on the dataset page: https://huggingface.co/datasets/Anwen0507/nlt-l16-l20-archive.