CoLoMo — ML Training Plugin for Claude Code

CoLoMo is a Claude Code plugin that provides ML training knowledge: Golden Rules for hyperparameter tuning, PyTorch code patterns, LLM fine-tuning guidance (LoRA/QLoRA/SFT/RLHF), and autonomous training agent workflows.

---

  # 1. 把这个 GitHub 仓库添加为 marketplace
  /plugin marketplace add Cola-Pig1121/CoLoMo
  # 2. 从该 marketplace 安装插件
  /plugin install colomo@Cola-Pig1121/CoLoMo

What This Plugin Provides

1. ML Training Skill (skills/ml-training/SKILL.md)

ML knowledge triggered automatically when discussing training topics:

• Golden Rules: Batch size formulas, learning rate linear scaling, optimizer selection by parameter count
• PyTorch Patterns: Training loops, gradient clipping, label smoothing, mixup, checkpointing, fine-tuning
• Distributed Training: DDP, DeepSpeed ZeRO stages, BF16/FP16 mixed precision, Flash Attention
• LLM Fine-tuning: 3-stage pipeline (pretrain → SFT → RLHF), LoRA/QLoRA configuration
• RAG: Chunking strategies, vector DB selection, hybrid retrieval

2. CoLoMo Agent (agents/colomo.md)

Autonomous ML training subagent — use it when the user wants to go from a requirement to a trained model.

Commands:

• /ml plan <requirement> — generate implementation plan
• /ml run — run training in Conda environment
• /ml test — run pytest in Conda environment
• /ml advise — GPU-based hyperparameter recommendations
• /ml explain <topic> — explain ML algorithm
• /ml rollback [n] — undo config changes

3. ML Rules (rules/ml/)

• coding-style.md: PyTorch conventions (device management, gradient handling, no in-place ops on pretrained weights, mixed precision)
• patterns.md: Golden Rules formulas, fine-tuning patterns (frozen backbone / differential LR / LoRA / QLoRA), augmentation, distributed training

---

Installation

Via Claude Code Marketplace (Recommended)

# 1. Add this GitHub repo as a marketplace
/plugin marketplace add Cola-Pig1121/CoLoMo

# 2. Install the plugin from that marketplace
/plugin install colomo@Cola-Pig1121/CoLoMo

Requires Claude Code v1.0.33+. Alternatively, open /plugin in Claude Code and browse the Marketplaces tab to add this repo.

---

Golden Rules Quick Reference

Signal	Rule	Action
CUDA OOM	BS = α × GPU_mem / (param_mem + activation_mem)	Halve batch_size
GPU util < 50%	—	Increase batch by 25%
Batch changed	LR_new = LR_old × (BS_new / BS_old)	Scale LR linearly
Params > 100M	—	Use AdamW
Params 10M–100M	—	Use Adam
Params < 10M	—	Use SGD
Batch reduced	—	Set grad_accum_steps = ceil(old / new)

Default α (safety_alpha) = 0.90 — reserve 10% VRAM headroom.

---

PyTorch Snippet Index

Snippet	Category	Use when
label_smoothing	Loss	Noisy labels, calibration
mixup	Augmentation	Generalization, adversarial robustness
grad_clip	Training	Deep transformers, RNNs
lr_decay	Scheduler	Cosine or step decay
checkpoint_save_load	IO	Resuming after interruption
finetune_fc	Finetune	Frozen backbone + train head
finetune_fc_high_lr_conv_low_lr	Finetune	Differential LR (BERT style)
no_weight_decay_bias	Optimizer	Standard fine-tuning practice
extract_imagenet_layer_feature	Feature	Transfer learning
train_visualization	Vis	Loss/accuracy plots

---

Architecture

CoLoMo/
├── .claude-plugin/
│   ├── plugin.json          # Plugin manifest
│   └── marketplace.json      # Marketplace discovery manifest
├── skills/
│   └── ml-training/
│       └── SKILL.md        # ML knowledge base (17 snippets + Golden Rules)
├── agents/
│   └── colomo.md           # CoLoMo subagent
├── rules/
│   └── ml/
│       ├── coding-style.md # PyTorch conventions
│       └── patterns.md      # Golden Rules + patterns
├── templates/
│   ├── pytorch-snippets/    # 17 standalone PyTorch snippets
│   ├── model_templates/     # Full project templates (pytorch-template)
│   └── docs/               # Algorithm references (LoRA, RAG, etc.)
├── docs/
│   ├── CONTRIBUTING.md     # Development guide
│   └── SETUP.md           # Configuration reference
└── CLAUDE.md              # Claude Code guidance

---

Fine-tuning Decision Tree

Is model > 1B params and single GPU?
├─ YES → QLoRA (4-bit NF4 base + LoRA adapters)
└─ NO
   ├─ Want maximum quality?
   │  ├─ YES → Full fine-tune or RLHF
   │  └─ NO
   │     └─ LoRA (r=8–16, single GPU, ~0.1–1% trainable params)
   └─ Have pretrained backbone?
      ├─ YES
      │  ├─ Small dataset → Frozen backbone + train head only
      │  └─ Medium dataset → Differential LR (high head / low backbone)
      └─ NO → Train from scratch

---

Contributing

See docs/CONTRIBUTING.md for development guidelines.