Guides

In-depth guides covering every aspect of the library.

• Environments

  ---

  Working with different environment types — single-turn, multi-turn, and tool-based.

  Environments

• Evaluation

  ---

  Running evaluations, computing metrics, and interpreting results.

  Evaluation

• Inference Backends

  ---

  Configure vLLM, HuggingFace, OpenAI, Anthropic, and OpenRouter backends.

  Backends

• Prompts

  ---

  Fragments, system prompts, templates, and model profiles.

  Prompts

• Tools & Function Calling

  ---

  Using tools in evaluations with auto-monitoring and MCP support.

  Tools

• Segmentation

  ---

  Multi-step reasoning with segmenters for chain-of-thought evaluation.

  Segmentation

• LLM-as-a-Judge

  ---

  Score open-ended responses with an LLM judge.

  Judge

• Dialogue

  ---

  LLM-in-the-loop environments for 20-questions, teacher, and custom setups.

  Dialogue

• Iterative Refinement

  ---

  Multi-turn refinement with code execution and judge feedback.

  Iterative

• Branching

  ---

  Checkpoint, branch, and explore environment states.

  Branching

• History Control

  ---

  Manage conversation history in multi-turn environments with windowing and reasoning stripping.

  History Control

• Parallelization

  ---

  Batched evaluation with lockstep trajectories and concurrent inference.

  Parallelization

• Containers

  ---

  Run environments in Docker or isolated subprocesses.

  Containers

• RL Training

  ---

  Integrate with veRL, TRL, and OpenRLHF for reward-based training.

  RL Training

• Multimodal Observations

  ---

  Image observations for vision-language models.

  Multimodal

• Library Landscape

  ---

  How llenvs compares to OpenEnv, verifiers, and similar libraries.

  Landscape