AGENT SKILL

modeling-and-layers

Understand Scenic model registry contracts, BaseModel task bases, Flax module construction, layers, attention, matchers, and tiny smoke checks.

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

01 前言 Front Matter

name
modeling-and-layers
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Understand Scenic model registry contracts, BaseModel task bases, Flax module construction, layers, attention, matchers, and tiny smoke checks.
disable-model-invocation
true

02 资源文件 3

03 SKILL.md 正文

Modeling and Layers

Use this sub-skill when the task is about Scenic model selection, BaseModel contracts, task-specific model bases, Flax module construction, layer/attention/matcher APIs, or tiny smoke checks.

Start here

Scope and routing

  • In scope: model registry lookups, BaseModel and task-base contracts, loss/metric normalization, Flax module construction, layer/attention/matcher APIs, and smoke-test patterns for tiny inputs.
  • Out of scope: training orchestration, launch commands, schedules, checkpoint lifecycle, dataset builders, project selection, and paper/baseline choice.
  • Route training/runtime execution to running-and-training.
  • Route dataset builders and input-pipeline work to data-pipelines.
  • Route project-level or paper-level model selection to baselines-and-projects.

Operating rules

  1. Begin with the registry; an unrecognized model name is usually a registry or config problem, not a training problem.
  2. Treat build_flax_model() as the construction boundary. Build the Flax module, then use tiny init/apply checks instead of a full run.
  3. Keep loss functions local and scalar; keep metrics cross-device safe by returning sum/normalizer pairs.
  4. Use tiny dummy inputs for smoke checks and shape assertions. Prefer init_with_output, init, and apply on minimal batches.
  5. Keep optional matcher dependencies explicit. Do not promise scipy or ott-jax behavior unless the target matcher requires it.

When to read what

Need Read
Registry, BaseModel, task bases, loss/metric rules, smoke-init pattern references/modeling-api.md
Attention, layer, masked-layer, and matcher APIs plus tiny validation patterns references/layers-and-matchers.md
Common failure modes and recovery references/troubleshooting.md
List registered models or resolve one class scripts/model_registry_probe.py

04 同仓库更多 Skills