InternLM/InternBootcamp

What it solves

InternAgentHarness (also known as InternBootcamp) provides a scalable synthetic environment framework for training and evaluating LLM agents. It moves beyond static benchmarks by creating executable, interactive, and verifiable task environments where models can learn from real tool calls, state feedback, and multi-round interactions.

How it works

The framework encapsulates agent tasks into trainable environment instances. Each task consists of four core components:

1. Instruction Generation: Creates task goals, initial states, and input contexts.
2. Tool Execution: Provides external tools that the agent can call and that can be verified.
3. Interaction Control: Manages the multi-round communication flow between the model and the environment.
4. Reward Calculation: Provides feedback signals based on the final result and intermediate processes.

It includes a BootCampCLI to automate the conversion of task configurations into interactive agent flows, enabling batch trajectory collection, SFT data filtering, RL rollout, and failure analysis.

Who it’s for

Researchers and developers who want to build agent training environments, evaluate a model's tool-use capabilities, or collect multi-round interaction trajectories to improve task design.

Highlights

• Multi-round Tool-calling: Specifically designed to handle complex reasoning tasks requiring asynchronous tool calls and state control within RL frameworks.
• Verifiable Environments: Includes diverse task examples such as battery design, robot trajectory planning, financial prediction, and Texas Hold'em.
• Distributed Tool Server: Features a Master-Worker architecture to support high-concurrency tool execution across multiple machines.
• Closed-loop Workflow: Integrates evaluation, data generation, and training into a single iterative cycle.