curry-train

curryTrain

A methodology-first deep learning training framework, packaged as a Claude Code plugin.

Idea is cheap. Infrastructure that lets you validate ideas fast is valuable.

中文 README → README.zh.md

What it is

curryTrain organizes deep learning training around the actual end-to-end workflow, not around an algorithm catalog. The plugin provides Skills, Agents, and a minimal Python template that scaffolds a new training project and assists you through six well-defined stages.

Six-stage methodology

Stage	Question it answers	Representative skills
1. Skeleton	Does the architecture exist and does data flow through it?	scaffolder, preflight-asserts, data-pipeline
2. Sanity	Is the implementation actually correct?	overfit-single-batch, init-loss-check, grad-flow-viz
3. Pre-validate	Will this idea pay off, before I burn the compute?	lr-range-test, small-scale-ablation, multi-seed-variance, mup-coord-check, scaling-fit, surrogate-task, compute-budget, kill-criterion
4. Scale-up	Will it scale stably to the target size?	capacity-sweep, optuna-integration, parallel-primitive-intro
5. Stabilize	Will it survive a long run?	warmup-cosine, loss-spike-rollback, checkpoint-cadence, run-journal
6. Iterate	Which experiment was actually better?	variance-aware-decision, error-cluster, ablation-matrix, runs-diff

Stage 3 is where most projects waste compute and where curryTrain provides the most differentiated value.

Components

• 47 skills organized as: 1 user-invoked (slash) + 4 workflow + 24 methodology + 14 primitive + 4 infra. Only /curry-train:init is exposed as a slash command; the other 46 skills auto-activate from natural-language phrasing in your messages.
• 4 specialized agents: scaffolder, hpo-proposer, failure-diagnoser, runs-diff
• 0 hooks, 0 MCP servers — the plugin stays light and explicit
• Python template at template/curry_train/ — a minimal layered scaffold (Runtime / Primitive / Model) you copy into your project via /curry-train:init

Installation

Option A — Claude Code marketplace (recommended)

In Claude Code, run:

/plugin marketplace add curryfromuestc/curry-train
/plugin install curry-train@curry-train

This adds the GitHub repo as a marketplace and installs the curry-train plugin from it. After installation, the /curry-train:init slash command and all description-activated skills (workflow, methodology, primitive, infra) become available in your sessions.

Option B — Local development install

If you cloned this repo locally and want to edit the plugin while using it:

git clone https://github.com/curryfromuestc/curry-train.git
mkdir -p ~/.claude/plugins
ln -s "$(pwd)/curry-train" ~/.claude/plugins/curry-train

Reload Claude Code (or run /reload-plugins) and the plugin will be picked up.

Option C — Per-session plugin dir

claude --plugin-dir /path/to/curry-train

Quick start

/curry-train:init is the only explicit slash command; everything else activates from natural-language phrasing.

# Bootstrap a new training project (copies the Python template into ./curry_train)
/curry-train:init my-experiment

Then drive the rest of the workflow by describing what you want:

• "scaffold a new model called my-model" → new-experiment skill (Stage 1)
• "smoke-test the runtime / time one optimizer step" → bench skill
• "this run crashed / loss went to NaN, help me debug" → diagnose skill
• "compare run A and run B / did this change actually help?" → runs-diff skill
• "how do I check my init loss is reasonable", "find a learning rate", "is my improvement real or noise" → the matching methodology skill auto-activates

This is by design: the methodology lives in skills and trips on what you describe, so you don't have to memorize a command surface.

Stack opinions (V1)

• Hydra + OmegaConf for config
• Lightning Fabric (not the Trainer) for distributed launch
• Optuna for hyperparameter search
• Logger protocol with TensorBoard as the default backend (no lock-in to W&B / MLflow)
• torchrun for launch (no custom launcher)

Philosophy

Architecture inspired by NVIDIA Bumblebee's three-layer split (Runtime ↔ Primitive ↔ Model). Workflow inspired by Karpathy's "A Recipe for Training Neural Networks". Built for engineers who train models — including unconventional ones (SNN, CV, multimodal) — and need fast, trustworthy iteration.

The framework intentionally keeps the Python core small. The framework's value is in methodology (skills), not in re-implementing what Lightning Fabric / Accelerate / DeepSpeed already do well.

Layout