0.6B
Datasets
All datasets matching “0.6B”UltraData-Math-L2-preview-embedded_selected_top20k_per_ncert_chapter_qwen3_0.6bTalkPlayData-Extra-attributes-qwen3_embedding_0.6bQwen3-0.6B-ExtractedUltraData-Math-L3-Textbook-Exercise-Synthetic-split-qwen3-0.6b-embeddedqwen3-0.6b-SAEOne JumpReLU SAE per layer of Qwen3-0.6B, trained by a scheduler that drives itself. All 28 layers are in.
Twenty-eight residual streams, twenty-eight sparse autoencoders, I didn't tweak any hyperparameters. Every SAE in this repo: d_in=1024, 32,768 features (32x expansion), JumpReLU activation, trained on streamed FineWeb-Edu at a target sparsity of L0=50. Sparsity was coerced, not calculated.
The run is finished. 1.81B tokens across 55,277 steps, every layer early-stopped on its own… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/qwen3-0.6b-SAE.audioset-dasheng-0.6b-emb
AudioSet DaSheng-0.6B embeddings
Mean-pooled, float16 embeddings of
danjacobellis/audioset_opus_24kbps
from mispeech/dasheng-0.6B.
Columns
path: source clip path (string)
label: source AudioSet label indices (list of int64)
emb: 1,280-dimensional fixed-size list of float16
Audio is decoded from the source Opus bytes, mixed to mono, and resampled to
16 kHz. The embedding is the model's documented outputdim=None output:
sigmoid applied to the mean of the final… See the full description on the dataset page: https://huggingface.co/datasets/quinnlue/audioset-dasheng-0.6b-emb.