tinker-agent

An intelligent agent for Anthropic's Tinker fine-tuning platform. Automates the creation of fine-tuning datasets and configuration with interactive and programmatic interfaces.

Installation

pip install tinker-agent

Or with uv:

uv pip install tinker-agent

Quick Start

Interactive Mode

Simply run the command without arguments for an interactive setup:

tinker-agent

This will guide you through:

1. Environment setup - Configure API keys on first run
2. Task selection - Choose between SFT, RL, or CPT
3. Model selection - Pick from available models
4. Dataset configuration - Specify your training data (see below)

Non-Interactive Mode

Pass configuration directly via command-line arguments:

tinker-agent \
  config.dataset=HuggingFaceFW/fineweb \
  config.task_type=sft \
  config.model=Qwen/Qwen3-8B

Dataset Options

You can train on either a HuggingFace dataset or a local directory:

# HuggingFace dataset (org/dataset-name format)
tinker-agent config.dataset=ServiceNow-AI/R1-Distill-SFT config.task_type=sft

# Local directory (e.g., your Obsidian vault, markdown notes, or custom data)
tinker-agent config.dataset=~/Documents/my-obsidian-vault config.task_type=sft
tinker-agent config.dataset=/path/to/training-data config.task_type=sft

Local directories are mounted as read-only - the agent can read your data but won't modify it. Supported file formats include .json, .jsonl, .parquet, .csv, .txt, and .md.

Configuration

Environment Setup

Run the setup command to configure your environment:

tinker-agent setup

This creates a .env file with:

• TINKER_API_KEY - Your Tinker API key (Get it here)
• WANDB_API_KEY - Weights & Biases API key for tracking (Get it here)
• WANDB_PROJECT - W&B project name for organizing experiments

Alternatively, set these as environment variables before running.

Task Types

• sft - Supervised Fine-Tuning (instruction-response pairs)
• rl - Reinforcement Learning (reward-based training)
• cpt - Continued Pre-Training (raw text data)

Available Models

Model	Type	Size
Qwen/Qwen3-VL-235B-A22B-Instruct	Vision	Large
Qwen/Qwen3-VL-30B-A3B-Instruct	Vision	Medium
Qwen/Qwen3-235B-A22B-Instruct-2507	Instruction	Large
Qwen/Qwen3-30B-A3B-Instruct-2507	Instruction	Medium
Qwen/Qwen3-30B-A3B	Hybrid	Medium
Qwen/Qwen3-30B-A3B-Base	Base	Medium
Qwen/Qwen3-32B	Hybrid	Medium
Qwen/Qwen3-8B	Hybrid	Small
Qwen/Qwen3-8B-Base	Base	Small
Qwen/Qwen3-4B-Instruct-2507	Instruction	Compact
openai/gpt-oss-120b	Reasoning	Medium
openai/gpt-oss-20b	Reasoning	Small
deepseek-ai/DeepSeek-V3.1	Hybrid	Large
deepseek-ai/DeepSeek-V3.1-Base	Base	Large
meta-llama/Llama-3.1-70B	Base	Large
meta-llama/Llama-3.3-70B-Instruct	Instruction	Large
meta-llama/Llama-3.1-8B	Base	Small
meta-llama/Llama-3.1-8B-Instruct	Instruction	Small
meta-llama/Llama-3.2-3B	Base	Compact
meta-llama/Llama-3.2-1B	Base	Compact
moonshotai/Kimi-K2-Thinking	Reasoning	Large

Usage Examples

Interactive Mode Example

$ tinker-agent

╭──────────────────────────────────────────╮
│            tinker-agent                  │
│     Fine-tuning configuration            │
╰──────────────────────────────────────────╯

Select task type:
  1    sft    Supervised Fine-Tuning
  2    rl     Reinforcement Learning
  3    cpt    Continued Pre-Training
Choice [1/2/3/sft/rl/cpt]: 1

Select model:
Key   Model                                Type         Size
1     Qwen/Qwen3-VL-235B-A22B-Instruct     Vision       Large
2     Qwen/Qwen3-VL-30B-A3B-Instruct       Vision       Medium
3     Qwen/Qwen3-235B-A22B-Instruct-2507   Instruction  Large
...
Choice (number or model name) [1]: 8

HuggingFace dataset: HuggingFaceFW/fineweb

Non-Interactive Example

# Basic usage
tinker-agent config.dataset=my-org/my-dataset config.task_type=sft

# With all options
tinker-agent \
  config.dataset=HuggingFaceFW/fineweb \
  config.task_type=sft \
  config.model=meta-llama/Llama-3.3-70B-Instruct

Environment Variables

# Set via environment
export TINKER_API_KEY="your-api-key"
export WANDB_API_KEY="your-wandb-key"
export WANDB_PROJECT="my-finetuning-project"

# Run with config
tinker-agent config.dataset=my-dataset config.task_type=rl

Features

• ✅ Interactive CLI - Beautiful rich terminal UI for configuration
• ✅ Non-interactive mode - Scriptable with command-line arguments
• ✅ Flexible datasets - Use HuggingFace datasets or local directories (Obsidian vaults, markdown notes, etc.)
• ✅ Model selection - Choose from available models
• ✅ Environment management - Simple .env-based configuration
• ✅ Sandboxed execution - Agent runs in isolated directory with path validation
• ✅ Trace viewer - Streamlit-based viewer for execution traces

Sandboxing

The agent runs in a sandboxed environment with strict path validation:

• Root directory isolation - Agent can only access files within its working directory
• Path validation - Blocks access to ~, $HOME, absolute paths, and .. escapes
• No system access - Cannot read sensitive files like /etc/passwd or user home directories

This ensures the agent operates safely without requiring Docker, making it more scalable for production use.

Additional Commands

View Execution Traces

tinker-viewer

Opens a Streamlit interface to view and analyze agent execution traces.

Development

Setup Development Environment

git clone https://github.com/anthropics/tinker-agent.git
cd tinker-agent
uv sync --extra dev

Run Tests

uv run pytest

Build Package

uv build

Deploy to PyPI

python deploy.py

This will:

1. Ask for confirmation
2. Request your PyPI API token (create one here)
3. Clean old builds
4. Build the package
5. Upload to PyPI

License

MIT

Contributing

Contributions welcome! Please open an issue or PR.