datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
attention-mechanism
The Attention Computation — Inside the Score
Companion dataset for Episode 13 of EXD: the attention mechanism.
One score matrix, one softmax, one weighted blend. This notebook loads
Qwen3.6-35B-A3B, runs it up to a real full-attention layer, and
reconstructs the entire attention computation by hand — QK Norm, RoPE,
QKᵀ, causal mask, softmax, weighted sum, gate, and output projection —
verified bit-for-bit against the model's own output.
📖 Article:… See the full description on the dataset page: https://huggingface.co/datasets/EXDai/attention-mechanism.SpaRRTa-Attention
SpaRRTa-Attention: Attention-Analysis Split of the SpaRRTa Benchmark
SpaRRTa-Attention is the interpretability asset for the synthetic
SpaRRTa benchmark. Each scene ships
with per-object segmentation masks so that a frozen Visual Foundation Model's self-attention
can be measured between the objects in the scene (Human / Tree / Truck), the CLS token,
the background, and register tokens.
📄 Paper: arXiv:2601.11729
💻 Code: github.com/gmum/SpaRRTa (see sparrta/analysis/)
🧩 Main (synthetic)… See the full description on the dataset page: https://huggingface.co/datasets/turhancan97/SpaRRTa-Attention.attention-uq-800q-colab
Attention/UQ 800-question Colab bundle
A deterministic 200-question subset for each of MultiModalQA, WebQA, HotpotQA, and TAT-QA. See manifest.json for exact upstream sources, hashes, counts, and the explicitly constructed WebQA distractor setting.
BDD-X_attentionboltzmann-attention-steering-artifactsface-attention-focus-on-eyearithmetic-attention-h100-results
Arithmetic Attention H100 Results
Run: 20260401_000525
Paper: Canavesi (2026), "Semiprime Bottlenecks in Arithmetic Graphs"
Configuration
GPU: H100 80GB
d_model: 128, heads: 4, layers: 4
Sequence lengths: up to 2048
Seeds per experiment: 2
Experiments
Mask structure and sparsity scaling
BFS routing efficiency (diameter analysis)
Distance-dependent accuracy (scaled)
Wall-clock speedup with Triton sparse kernel
Gradient-based position importance… See the full description on the dataset page: https://huggingface.co/datasets/thotik/arithmetic-attention-h100-results.Attention_Diffkeras-attention