v0.1.2

What's changed

With the v0.1.2 release, verifiers is significantly more production-ready, and stable to build and train with. We appreciate everyone's patience with the changes and bug fixes thus far as we've addressed a number of long-time requests, and are excited to see what you all build with it!

Highlights:

• Proper encapsulation of Environments as standalone modules (see environments/), which can contain their own dependencies in a pyproject.toml, and need only to expose a load_environment(...) -> vf.Environment function in order to be trainable.
• Script flows for initializing (vf-init), installing (vf-install), and evaluating (vf-eval) Environments before training.
• Reorganization of examples and training scripts, removing lots of duplicated logic and creating a cleaner separation between library code and example code.
• Deprecation of the manual dynamically-batched LLM inference worker in favor of proper AsyncLLM support, allowing full control of native vLLM sampling parameters.
• Support for native tool call parsing + parallel tool calls in ToolEnv (replacing the manual XMLParser approach).
• Another trainer! Environments built with verifiers are now trainable with prime-rl (as of 58ac91f for v0.1.2), which supports multi-node FSDP async training, is the primary RL framework used by the Prime Intellect research team, and is under ongoing development and stress-testing in advance of large-scale multi-environment training runs.
• Pydantic types for core data classes used by Environments.
• Improvements to GRPOTrainer, including supporting a single max_seq_len option (instead of separate prompt + completion lengths), and configurable turn length limits via max_tokens.
• Many more Environment examples.
• Improved logging and evaluation options.
• Overhauled README.md and docs.