AGENT SKILL

attention-views

Route BertViz head_view and model_view attention-visualization tasks.

License 许可
Apache-2.0
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
attention-views
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Route BertViz head_view and model_view attention-visualization tasks.
disable-model-invocation
true

02 资源文件 3

03 SKILL.md 正文

attention-views

Use this sub-skill for BertViz attention visualizations built from bertviz.head_view.head_view, bertviz.model_view.model_view, and bertviz.util.format_attention.

Route here when

  • You need self-attention, sentence-pair, or encoder-decoder/cross-attention visualizations.
  • You need html_action='return' for saved HTML or non-notebook workflows.
  • You need include_layers, include_heads, layer, heads, sentence_b_start, display_mode, or token prettification control.
  • You want a quick offline smoke check using synthetic tensors.

Do not route here when

  • The task is about query/key neuron inspection or get_attention; use ../neuron-view/SKILL.md.
  • The task is generic Hugging Face training, serving, or model selection.
  • The task needs vendored JS assets; BertViz loads its own package data.

Start with

Routing shortcuts

  • Self-attention: pass attention plus tokens.
  • Sentence pairs: add sentence_b_start to split the token list.
  • Encoder-decoder: pass encoder_attention, decoder_attention, cross_attention, encoder_tokens, and decoder_tokens.
  • head_view uses layer and heads as default selections.
  • model_view uses include_layers and include_heads as filters, plus display_mode='dark'|'light'.
  • For notebooks and scripts that save HTML, call with html_action='return' and write the returned .data.
  • Use include_layers and include_heads to keep long inputs responsive.

04 同仓库更多 Skills