Official Resources

tinker-atropos

Standalone Atropos integration with Thinking Machines Tinker API. RL training infrastructure for fine-tuning tool-calling models on real agent trajectories.

Upstream ↗Seed list ↗Role: tooling
Why it matters

Profile

Standalone Atropos integration with Thinking Machines Tinker API. RL training infrastructure for fine-tuning tool-calling models on real agent trajectories.

setup mediumintegration highinterface api
Provenance

Signals

Listed in the awesome-hermes-agent README

Sources: 2 / Surfaces: 1

Fast skim

What the upstream surface says

Short excerpt only, so you can decide whether to click out.

An integration layer connecting Atropos (https://github.com/NousResearch/atropos) with the Thinking Machines Tinker API (https://thinkingmachines.ai/tinker/). This package enables seamless model training with Atropos environments from your local machine, abstracting away compute management and infrastructure concerns.

First, obtain a Tinker API key from https://tinker-console.thinkingmachines.ai/keys.

Run the following commands in separate terminal windows to start a training run:

tinker-atroposInstallationQuickstartTerminal 1: Start Atropos APITerminal 2: Start trainingTerminal 3: Start environment (use built-in or any Atropos environment)Using Any Atropos EnvironmentUse any Atropos environment with Tinker training
  • The env section for environment configuration
  • The openai section for inference server configuration
  • The tinker section is used by the trainer (ignored by the environment)
  • gsm8k_tinker.py - Math reasoning with GSM8k dataset
  • default.yaml - Standard GSM8k config (50 steps, batch_size=128, Llama-3.1-8B)
  • quick_test.yaml - Quick test (10 steps, Llama-3.1-8B, no wandb)
  • model_name: "meta-llama/Llama-3.1-8B-Instruct"
  • Add a tinker section with training parameters: