Together AI Fine-Tuning Skill for Claude Code

What it is

This is a Claude Code skill that teaches the AI agent how to use Together AI's fine-tuning API. It covers LoRA (efficient adapter tuning), full fine-tuning, DPO preference tuning, VLM (vision-language model) training, and function-calling fine-tuning.

The skill is designed for developers who want their Claude Code agent to handle fine-tuning tasks without repeatedly explaining Together AI's API surface. Install the skill once and the agent knows how to create training jobs, format datasets, and monitor runs.

How it saves time or tokens

Without this skill, every fine-tuning conversation starts with pasting Together AI docs into the chat. The skill pre-loads structured knowledge about API endpoints, dataset formats, hyperparameter defaults, and common workflows. This reduces per-conversation token usage by an estimated 2,700 tokens.

The skill also encodes best practices: when to use LoRA vs full fine-tuning, how to structure DPO preference pairs, and what hyperparameters to start with for different model sizes.

How to use

1. Install the skill:

npx skills add togethercomputer/skills

1. The skill is now available in your Claude Code sessions. Ask Claude Code to fine-tune a model:

You: Fine-tune Llama 3 on my customer support dataset using LoRA

Claude: I will use Together AI's fine-tuning API with LoRA.
Here is the configuration...

1. Claude Code generates the correct API calls, dataset formatting, and monitoring commands.

Example

# Together AI fine-tuning API call (generated by the skill)
import together

client = together.Together(api_key='your-key')

# Create a LoRA fine-tuning job
response = client.fine_tuning.create(
    training_file='file-abc123',
    model='meta-llama/Llama-3-8b',
    n_epochs=3,
    learning_rate=1e-5,
    lora=True,
    lora_r=16,
    lora_alpha=32
)

print(f'Job ID: {response.id}')
print(f'Status: {response.status}')

Related on TokRepo

• AI Tools for Coding -- Other AI coding tools and skills for development workflows
• Prompt Library -- Curated prompt templates and skills for AI agents

Common pitfalls

• Together AI requires an API key with billing enabled. Fine-tuning jobs incur compute costs that vary by model size and training duration.
• Dataset format matters: chat fine-tuning expects JSONL with messages arrays, while completion fine-tuning expects prompt/completion pairs. The skill handles this distinction.
• LoRA is the right default for most use cases. Full fine-tuning is only necessary when you need to change the model's behavior fundamentally, and it costs significantly more.