custom-model-planning

Plugin custom-model-planning in danielrosehill/Claude-Custom-Model-Planning-Plugin. Commands read from the repo's manifest — 6 days ago.

Skills (20)

• capture-model-idea

  Write down an idea for a custom model, fine-tune or classifier as a short idea note, before deciding whether it earns a repo.

• compare-approaches

  Should I fine-tune, use an existing model, prompt an LLM or skip the model? Compare build options for a model idea and record a verdict.

• create-training-run

  Create and launch a training run from a saved recipe on Modal, RunPod or local Unsloth. Use for "train with this recipe", "new recipe".

• dataset-map

  Show and keep current the hub's map of shared training datasets, meaning one canonical corpus, the views cut from it, and which projects train on each. Use before planning training data for a project that reuses existing recordings or rows, when adding a new view or consumer, when asked "which datas

• evaluate-incremental-training

  Decide whether a model can be updated incrementally from its last checkpoint on the data added since, or needs training from scratch. Goes beyond counting new data: checks that the new rows are compatible with the parent run (tokenizer, labels, audio or input format, schema, splits, distribution, re

• finish-training-run

  Close out a model project's training runs once they are done — pick the run to release, publish it to Hugging Face with a model card, archive its weights for safekeeping, verify both copies, then delete every weight, checkpoint and leftover the project's runs left on the training compute's storage (

• frame-the-task

  Turn a model idea into an ML task: task type, I/O contract, labels, constraints and success metric, as numbered requirements.

• model-exists-check

  Is there already a model that does this? Quick Hugging Face Hub triage of models, fine-tunes, datasets and papers for a task idea.

• model-project-status

  Where is this model project up to? Report stage, verdict and next step for one planning repo, or every project in the hub registry.

• plan-evaluation

  Define how a planned model is judged: metrics, eval set, a mandatory no-model baseline and the pass bar, before any training starts.

• plan-training-data

  Plan training and eval data for a fine-tune or classifier: sources, labelling, synthetic generation, size estimates, splits and licences.

• run-training-job

  Run a training, fine-tuning, sweep or GPU evaluation job for a model project on the user's designated training compute (config `training_compute`, e.g. a general-purpose Modal training app) rather than an ad-hoc or project-unrelated app. Use when the user says "train it", "kick off a training run",

• select-base-model

  Pick a base model to fine-tune: shortlist small open-weights candidates by size, licence, language coverage and task fit, then compare.

• setup-cloud-training

  Install and authenticate the Modal CLI and RunPod CLI for cloud GPU training. Use for "set up Modal", "install runpodctl".

• setup-recipes-repo

  Create or connect the private GitHub repo that stores training recipes and save its location. Use for "set up my recipes repo".

• setup-unsloth-local

  Install Unsloth for on-device fine-tuning on a local NVIDIA GPU, after checking the hardware. Use for "set up Unsloth", "train locally".

• standardize-dataset

  Align a dataset built up over time, and every subset it draws on, to one convention before training from scratch. Works backwards from a run that has to start again anyway: inventory the conventions actually in use, write one standard, record each decision, dry-run, rewrite labels with the old value

• start-model-project

  Create the project repo for a custom model or fine-tune idea (private by default; a public lifecycle repo when the model will be released publicly), seed its docs and project.json, and register it in the hub.

• survey-prior-art

  Deep prior-art research for a model idea: Hub models, fine-tunes, datasets, papers, Spaces and GitHub, written up as research findings.

• write-model-spec

  Write the build spec for a custom model or fine-tune from the planning docs, plus a draft Hugging Face model card.

Subagents (1)

• prior-art-scout

  Searches one source class (Hub models, datasets, papers, or code and demos) for prior art on a framed ML task; returns JSON findings.

Similar plugins