base on Verifiers for LLM Reinforcement Learning <p align="center"> </p> <p align="center"> <img src="https://github.com/user-attachments/assets/40c36e38-c5bd-4c5a-9cb3-f7b902cd155d#gh-light-mode-only" alt="Prime Intellect" width="312"> <img src="https://github.com/user-attachments/assets/6414bc9b-126b-41ca-9307-9e982430cde8#gh-dark-mode-only" alt="Prime Intellect" width="312"> </p> --- <h3 align="center"> Verifiers: Environments for LLM Reinforcement Learning </h3> <!--- <p align="center"> Created by Will Brown (<a href="https://github.com/willccbb">@willccbb</a>) </p> ---> --- <p align="center"> <a href="https://github.com/willccbb/verifiers/actions/workflows/style.yml"> <img src="https://github.com/willccbb/verifiers/actions/workflows/style.yml/badge.svg" alt="Style" /> </a> <a href="https://github.com/willccbb/verifiers/actions/workflows/test.yml"> <img src="https://github.com/willccbb/verifiers/actions/workflows/test.yml/badge.svg" alt="Test" /> </a> <a href="https://github.com/willccbb/verifiers/actions/workflows/publish-environments.yml"> <img src="https://github.com/willccbb/verifiers/actions/workflows/publish-environments.yml/badge.svg" alt="Envs" /> </a> </p> ## Overview Verifiers is a library of modular components for creating RL environments and training LLM agents. Environments built with Verifiers can be used directly as LLM evaluations, synthetic data pipelines, or agent harnesses for any OpenAI-compatible model endpoint, in addition to RL training. Verifiers includes an async GRPO implementation built around the `transformers` Trainer, is supported by `prime-rl` for large-scale FSDP training, and can easily be integrated into any RL framework which exposes an OpenAI-compatible inference client. Full documentation is available [here](https://verifiers.readthedocs.io/en/latest/). ## Setup We recommend using `verifiers` with along [uv](https://docs.astral.sh/uv/getting-started/installation/) for dependency management in your own project: ```bash # install uv (first time only) curl -LsSf https://astral.sh/uv/install.sh | sh # create a fresh project -- 3.11 + 3.12 supported uv init && uv venv --python 3.12 source .venv/bin/activate ``` For local (CPU) development and evaluation with API models, do: ```bash uv add verifiers # uv add 'verifiers[dev]' for Jupyter + testing support ``` For training on GPUs with `vf.GRPOTrainer`, do: ```bash uv add 'verifiers[all]' && uv pip install flash-attn --no-build-isolation ``` To use the latest `main` branch, do: ```bash uv add verifiers @ git+https://github.com/willccbb/verifiers.git ``` To use with `prime-rl`, see [here](https://github.com/PrimeIntellect-ai/prime-rl). To install `verifiers` from source for core library development, do: ```bash git clone https://github.com/willccbb/verifiers.git cd verifiers # for CPU-only dev: uv sync --extra dev # or, for trainer dev: uv sync --all-extras && uv pip install flash-attn --no-build-isolation # install pre-commit hooks uv run pre-commit install ``` In general, we recommend that you build and train Environments *with* `verifiers`, not *in* `verifiers`. If you find yourself needing to clone and modify the core library in order to implement key functionality for your project, we'd love for you to open an issue so that we can try and streamline the development experience. Our aim is for `verifiers` to be a reliable toolkit to build on top of, and to minimize the "fork proliferation" which often pervades the RL infrastructure ecosystem. ## Environments Environments in Verifiers are installable Python modules which can specify dependencies in a `pyproject.toml`, and which expose a `load_environment` function for instantiation by downstream applications (e.g. trainers). See `environments/` for examples. To initialize a blank Environment module template, do: ```bash vf-init vf-environment-name # -p /path/to/environments (defaults to "./environments") ``` To an install an Environment module into your project, do: ```bash vf-install vf-environment-name # -p /path/to/environments (defaults to "./environments") ``` To install an Environment module from this repo's `environments` folder, do: ```bash vf-install vf-math-python --from-repo # -b branch_or_commit (defaults to "main") ``` Once an Environment module is installed, you can create an instance of the Environment using `load_environment`, passing any necessary args: ```python import verifiers as vf vf_env = vf.load_environment("vf-environment-name", **env_args) ``` To run a quick evaluation of your Environment with an API-based model, do: ```bash vf-eval vf-environment-name -s # run and save eval results locally # vf-eval -h for config options; defaults to gpt-4.1-mini, 5 prompts, 3 rollouts for each ``` The core elements of Environments are: - Datasets: a Hugging Face `Dataset` with a `prompt` column for inputs, and optionally `answer (str)` or `info (dict)` columns for evaluation (both can be omitted for environments that evaluate based solely on completion quality) - Rollout logic: interactions between models and the environment (e.g. `env_response` + `is_completed` for any `MultiTurnEnv`) - Rubrics: an encapsulation for one or more reward functions - Parsers: optional; an encapsulation for reusable parsing logic We support both `/v1/chat/completions`-style and `/v1/completions`-style inference via OpenAI clients, though we generally recommend `/v1/chat/completions`-style inference for the vast majority of applications. Both the included `GRPOTrainer` as well as `prime-rl` support the full set of [SamplingParams](https://docs.vllm.ai/en/v0.6.0/dev/sampling_params.html) exposed by vLLM (via their OpenAI-compatible [server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) interface), and leveraging this will often be the appropriate way to implement rollout strategies requiring finer-grained control, such as interrupting and resuming generations for interleaved tool use, or enforcing reasoning budgets. The primary constraint we impose on rollout logic is that token sequences must be *increasing*, i.e. once a token has been added to a model's context in a rollout, it must remain as the rollout progresses. Note that this causes issues with some popular reasoning models such as the Qwen3 and DeepSeek-R1-Distill series; see [Footguns](#footguns) for guidance on adapting these models to support multi-turn rollouts. ### SingleTurnEnv For tasks requiring only a single response from a model for each prompt, you can use `SingleTurnEnv` directly by specifying a Dataset and a Rubric. Rubrics are sets of reward functions, which can be either sync or async. ```python from datasets import load_dataset import verifiers as vf dataset = load_dataset("my-account/my-dataset", split="train") def reward_A(prompt, completion, info) -> float: # reward fn, e.g. correctness ... def reward_B(parser, completion) -> float: # auxiliary reward fn, e.g. format ... async def metric(completion) -> float: # non-reward metric, e.g. proper noun count ... rubric = vf.Rubric(funcs=[reward_A, reward_B, metric], weights=[1.0, 0.5, 0.0]) vf_env = SingleTurnEnv( dataset=dataset, rubric=rubric ) results = vf_env.evaluate(client=OpenAI(), model="gpt-4.1-mini", num_examples=100, rollouts_per_example=1) vf_env.make_dataset(results) # HF dataset format ``` Datasets should be formatted with columns for: - `'prompt' (List[ChatMessage])` OR `'question' (str)` fields - `ChatMessage` = e.g. `{'role': 'user', 'content': '...'}` - if `question` is set instead of `prompt`, you can also pass `system_prompt (str)` and/or `few_shot (List[ChatMessage])` - `answer (str)` AND/OR `info (dict)` (both optional, can be omitted entirely) - `task (str)`: optional, used by `EnvGroup` and `RubricGroup` for orchestrating composition of Environments and Rubrics The following named attributes available for use by reward functions in your Rubric: - `prompt`: sequence of input messages - `completion`: sequence of messages generated during rollout by model and Environment - `answer`: primary answer column, optional (defaults to empty string if omitted) - `state`: can be modified during rollout to accumulate any metadata (`state['responses']` includes full OpenAI response objects by default) - `info`: auxiliary info needed for reward computation (e.g. test cases), optional (defaults to empty dict if omitted) - `task`: tag for task type (used by `EnvGroup` and `RubricGroup`) - `parser`: the parser object declared. Note: `vf.Parser().get_format_reward_func()` is a no-op (always 1.0); use `vf.ThinkParser` or a custom parser if you want a real format adherence reward. **Note**: Some environments can fully evaluate using only `prompt`, `completion`, and `state` without requiring ground truth `answer` or `info` data. Examples include format compliance checking, completion quality assessment, or length-based rewards. For tasks involving LLM judges, you may wish to use `vf.JudgeRubric()` for managing requests to auxiliary models. Note on concurrency: environment APIs accept `max_concurrent` to control parallel rollouts. The `vf-eval` CLI currently exposes `--max-concurrent-requests`; ensure this maps to your environment’s concurrency as expected. `vf-eval` also supports specifying `sampling_args` as a JSON object, which is sent to the vLLM inference engine: ```bash vf-eval vf-environment-name --sampling-args '{"reasoning_effort": "low"}' ``` Use `vf-eval -s` to save outputs as dataset-formatted JSON, and view all locally-saved eval results with `vf-tui`. ### ToolEnv For many applications involving tool use, you can use `ToolEnv` to leverage models' native tool/function-calling capabilities in an agentic loop. Tools can be specified as generic Python functions (with type hints and docstrings), which will then be passed in JSON schema form to each inference request. ```python import verifiers as vf vf_env = vf.ToolEnv( dataset= ... # HF Dataset with 'prompt'/'question' and optionally 'answer'/'info' columns rubric= ... # Rubric object; vf.ToolRubric() can be optionally used for counting tool invocations in each rollout tools=[search_tool, read_article_tool, python_tool], # python functions with type hints + docstrings max_turns=10 ) ``` In cases where your tools require heavy computational resources, we recommend hosting your tools as standalone servers (e.g. MCP servers) and creating lightweight wrapper functions to pass to `ToolEnv`. Parallel tool call support is enabled by default. For training, or self-hosted endpoints, you'll want to enable auto tool choice in [vLLM](https://docs.vllm.ai/en/stable/features/tool_calling.html#automatic-function-calling) with the appropriate parser. If your model does not support native tool calling, you may find the `XMLParser` abstraction useful for rolling your own tool call parsing on top of `MultiTurnEnv`; see `environments/xml_tool_env` for an example. ### MultiTurnEnv Both `SingleTurnEnv` and `ToolEnv` are instances of `MultiTurnEnv`, which exposes an interface for writing custom Environment interaction protocols. The two methods you must override are ```python from typing import Tuple import verifiers as vf from verifiers.types import Messages, State class YourMultiTurnEnv(vf.MultiTurnEnv): def __init__(self, dataset: Dataset, rubric: Rubric, max_turns: int, **kwargs): async def is_completed(self, messages: Messages, state: State, **kwargs) -> bool: # return whether or not a rollout is completed async def env_response(self, messages: Messages, state: State, **kwargs) -> Tuple[Messages, State]: # return new environment message(s) + updated state ``` If your application requires more fine-grained control than is allowed by `MultiTurnEnv`, you may want to inherit from the base `Environment` functionality directly and override the `rollout` method. ## Training ### GRPOTrainer The included trainer (`vf.GRPOTrainer`) supports running GRPO-style RL training via Accelerate/DeepSpeed, and uses vLLM for inference. It supports both full-parameter finetuning, and is optimized for efficiently training dense transformer models on 2-16 GPUs. ```bash # install environment vf-install vf-wordle (-p /path/to/environments | --from-repo) # quick eval vf-eval vf-wordle -m (model_name in configs/endpoints.py) -n NUM_EXAMPLES -r ROLLOUTS_PER_EXAMPLE # inference (shell 0) CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 vf-vllm --model willcb/Qwen3-1.7B-Wordle \ --data-parallel-size 7 --enforce-eager --disable-log-requests # training (shell 1) CUDA_VISIBLE_DEVICES=6,7 accelerate launch --num-processes 2 \ --config-file configs/zero3.yaml examples/grpo/train_wordle.py --size 1.7B ``` Alternatively, you can train environments with the external `prime-rl` project (FSDP-first orchestration). See the `prime-rl` README for installation and examples. For example: ```toml # orchestrator config (prime-rl) [environment] id = "vf-math-python" # or your environment ID ``` ```bash # run (prime-rl) uv run rl \ --trainer @ configs/your_exp/train.toml \ --orchestrator @ configs/your_exp/orch.toml \ --inference @ configs/your_exp/infer.toml ``` ### Troubleshooting - Ensure your `wandb` and `huggingface-cli` logins are set up (or set `report_to=None` in `training_args`). You should also have something set as your `OPENAI_API_KEY` in your environment (can be a dummy key for vLLM). - If using high max concurrency, increase the number of allowed open sockets (e.g. `ulimit -n 4096`) - On some setups, inter-GPU communication can [hang](https://github.com/huggingface/trl/issues/2923) or crash during vLLM weight syncing. This can usually be alleviated by setting (or unsetting) `NCCL_P2P_DISABLE=1` in your environment (or potentially `NCCL_CUMEM_ENABLE=1`). Try this as your first step if you experience NCCL-related issues. - If problems persist, please open an [issue](https://github.com/willccbb/verifiers/issues). ### Resource Requirements `GRPOTrainer` is optimized for setups with at least 2 GPUs, scaling up to multiple nodes. 2-GPU setups with sufficient memory to enable small-scale experimentation can be [rented](https://app.primeintellect.ai/dashboard/create-cluster?image=ubuntu_22_cuda_12) for <$1/hr. ### PRIME-RL If you do not require LoRA support, you may want to use the `prime-rl` trainer, which natively supports Environments created using `verifiers`, is more optimized for performance and scalability via FSDP, includes a broader set of configuration options and user experience features, and has more battle-tested defaults. Both trainers support asynchronous rollouts, and use a one-step off-policy delay by default for overlapping training and inference. See the `prime-rl` [docs](https://github.com/PrimeIntellect-ai/prime-rl) for usage instructions. ## Further Documentation See the full [docs](https://verifiers.readthedocs.io/en/latest/) for more information. ## Contributions Verifiers warmly welcomes community contributions! Please open an issue or PR if you encounter bugs or other pain points during your development, or start a discussion for more open-ended questions. Please note that the core `verifiers/` library is intended to be a relatively lightweight set of reusable components rather than an exhaustive catalog of RL environments. For *applications* of `verifiers` (e.g. "an Environment for XYZ task"), you are welcome to submit a PR for a self-contained module that lives within `environments/` if it serves as a canonical example of a new pattern. Stay tuned for more info shortly about our plans for supporting community Environment contributions 🙂 ## Citation If you use this code in your research, please cite: ```bibtex @misc{brown_verifiers_2025, author = {William Brown}, title = {{Verifiers}: Reinforcement Learning with LLMs in Verifiable Environments}, howpublished = {\url{https://github.com/willccbb/verifiers}}, note = {Commit abcdefg • accessed DD Mon YYYY}, year = {2025} } ``` ## Roadmap - Default patterns for hosted resources such as code sandboxes, auxiliary models, and MCP servers - Multimodal input support - Non-increasing token sequences via REINFORCE ", Assign "at most 3 tags" to the expected json: {"id":"13882","tags":[]} "only from the tags list I provide: [{"id":77,"name":"3d"},{"id":89,"name":"agent"},{"id":17,"name":"ai"},{"id":54,"name":"algorithm"},{"id":24,"name":"api"},{"id":44,"name":"authentication"},{"id":3,"name":"aws"},{"id":27,"name":"backend"},{"id":60,"name":"benchmark"},{"id":72,"name":"best-practices"},{"id":39,"name":"bitcoin"},{"id":37,"name":"blockchain"},{"id":1,"name":"blog"},{"id":45,"name":"bundler"},{"id":58,"name":"cache"},{"id":21,"name":"chat"},{"id":49,"name":"cicd"},{"id":4,"name":"cli"},{"id":64,"name":"cloud-native"},{"id":48,"name":"cms"},{"id":61,"name":"compiler"},{"id":68,"name":"containerization"},{"id":92,"name":"crm"},{"id":34,"name":"data"},{"id":47,"name":"database"},{"id":8,"name":"declarative-gui "},{"id":9,"name":"deploy-tool"},{"id":53,"name":"desktop-app"},{"id":6,"name":"dev-exp-lib"},{"id":59,"name":"dev-tool"},{"id":13,"name":"ecommerce"},{"id":26,"name":"editor"},{"id":66,"name":"emulator"},{"id":62,"name":"filesystem"},{"id":80,"name":"finance"},{"id":15,"name":"firmware"},{"id":73,"name":"for-fun"},{"id":2,"name":"framework"},{"id":11,"name":"frontend"},{"id":22,"name":"game"},{"id":81,"name":"game-engine "},{"id":23,"name":"graphql"},{"id":84,"name":"gui"},{"id":91,"name":"http"},{"id":5,"name":"http-client"},{"id":51,"name":"iac"},{"id":30,"name":"ide"},{"id":78,"name":"iot"},{"id":40,"name":"json"},{"id":83,"name":"julian"},{"id":38,"name":"k8s"},{"id":31,"name":"language"},{"id":10,"name":"learning-resource"},{"id":33,"name":"lib"},{"id":41,"name":"linter"},{"id":28,"name":"lms"},{"id":16,"name":"logging"},{"id":76,"name":"low-code"},{"id":90,"name":"message-queue"},{"id":42,"name":"mobile-app"},{"id":18,"name":"monitoring"},{"id":36,"name":"networking"},{"id":7,"name":"node-version"},{"id":55,"name":"nosql"},{"id":57,"name":"observability"},{"id":46,"name":"orm"},{"id":52,"name":"os"},{"id":14,"name":"parser"},{"id":74,"name":"react"},{"id":82,"name":"real-time"},{"id":56,"name":"robot"},{"id":65,"name":"runtime"},{"id":32,"name":"sdk"},{"id":71,"name":"search"},{"id":63,"name":"secrets"},{"id":25,"name":"security"},{"id":85,"name":"server"},{"id":86,"name":"serverless"},{"id":70,"name":"storage"},{"id":75,"name":"system-design"},{"id":79,"name":"terminal"},{"id":29,"name":"testing"},{"id":12,"name":"ui"},{"id":50,"name":"ux"},{"id":88,"name":"video"},{"id":20,"name":"web-app"},{"id":35,"name":"web-server"},{"id":43,"name":"webassembly"},{"id":69,"name":"workflow"},{"id":87,"name":"yaml"}]" returns me the "expected json"