Verifiers Adapter¶

Wraps verifiers environments as llenvs MDP environments.

Installation¶

uv pip install -e ".[verifiers]"

Environment Types¶

The adapter supports two verifiers environment types:

verifiers Type	llenvs Wrapper	Turn Type
`SingleTurnEnv`	`VerifiersSingleTurnEnvironment`	Single-turn
`ToolEnv` / `MultiTurnEnv`	`VerifiersToolEnvironment`	Multi-turn with tools
`SandboxEnv` / `PythonEnv`	Not supported	Requires Prime Sandboxes

Quick Start¶

Single-Turn (GSM8K, MATH, etc.)¶

from llenvs.adapters.verifiers import VerifiersAdapter

adapter = VerifiersAdapter()
env = adapter.get_environment("gsm8k")

state, info = env.reset(options={"task_index": 0})
print(state.observation.prompt)  # Math question
print(info["system_prompt"])     # From verifiers env

from llenvs.core import Action
result = env.step(state, Action(text="The answer is \\boxed{42}"))
print(result.rewards.total)  # Rubric-computed reward

Tool Environment¶

env = adapter.get_environment("tool_test")

state, info = env.reset(options={"task_index": 0})
print(env.available_tools)  # Tool definitions from verifiers

# Use tools
from llenvs.core.tools import ToolCall
call = ToolCall(id="c1", name="tool_A", arguments={"x": 5})
action = Action(text="", tool_calls=(call,))
result = env.step(state, action)

With Evaluation Runner¶

from llenvs.evaluation.runner import run_evaluation

result = run_evaluation(
    environment=env,
    backend=backend,
    task_indices=list(range(100)),
)
print(f"Success rate: {result.success_rate}")

How It Works¶

Dataset Mapping¶

verifiers datasets are HuggingFace Dataset objects with columns:

prompt: List of chat message dicts (system + user messages)
answer: Ground truth string
example_id: Integer index

The adapter extracts the user message as the observation prompt, stores the full row in hidden state for reward computation.

Rubric Rewards¶

verifiers rubrics contain weighted async reward functions. The adapter:

Calls each rubric function with appropriate keyword arguments (completion, answer, prompt, etc.)
Applies the rubric's weights to each score
Returns the weighted sum as a single OUTCOME-type Signal named "verifiers_rubric"

# Rubric with two functions at different weights
rubric = vf.Rubric(
    funcs=[exact_match, format_check],
    weights=[2.0, 0.5],
)
# -> Signal(name="verifiers_rubric", reward_type=OUTCOME, reward=2.0*1.0 + 0.5*0.8)

Tool Conversion¶

For ToolEnv, the adapter converts OpenAI-format tool schemas to llenvs ToolDefinition objects and wraps the original Python callables in a VerifiersToolExecutor.

Parameters¶