Bounty verifiers

[banga-agents]

PR Type

• RL Environment PR - Complete Environment Snapshot & Zero-Training sections

---

📝 General Information

Description

Complete implementation of the VerifiersEnv adapter for Prime Intellect's Environment Hub, enabling Atropos to run environments installed via the verifiers library and Prime CLI.

Key Features:

• Full adapter with collect_trajectories(), score(), evaluate(), wandb_log()
• RubricGroup support for multi-rubric environments (like wordle)
• Optional dependency - clean ImportError if verifiers not installed
• 26 unit tests - CI passes without Prime credentials
• Comprehensive documentation in environments/README.md

This PR completes and extends the work started in #258.

---

🔖 Environment Snapshot

Field	Your Entry
Environment Name	VerifiersEnv (Prime Intellect Hub Adapter)
Short Description	Adapter to run Prime Intellect Environment Hub environments in Atropos via the verifiers library.
Category	Verifiable-Reasoning
Dataset Needed?	No (datasets provided by Prime Hub environments)
External Deps	verifiers (optional), Prime CLI (uv tool install prime)
Environmental Variables	OPENAI_API_KEY (for LLM inference)
Compute Footprint Estimate	<1 GB RAM, depends on loaded environment

🧪 Zero-Training Test Results

Details

W&B Link: N/A (tested locally)

Verification Test with will/wordle environment:

✓ rubric type: RubricGroup
✓ has rubrics attr: True
✓ num rubrics: 2
✓ reward_funcs: ['__score_rollout__']
✓ using score_rollout: True
✓ Training items: 2000
✓ Eval items: 20

Unit Test Results:

======================== 26 passed, 3 skipped in 7.41s =========================

Examples of the Environment scoring:

The adapter successfully:

1. Loads environments from Prime Hub via vf.load_environment()
2. Extracts rubrics (single or RubricGroup)
3. Normalizes datasets to Atropos format
4. Provides scoring via score_rollout() or reward functions

---

✅ Developer & Reviewer Checklist

• Code follows project style (black, isort, flake8 pass with pre-commit)
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• New and existing unit tests pass locally with my changes
• Docstrings added for all new public classes / functions
• If .env vars required, did you add it to the .env.example in repo root?

---

Verification Steps for Maintainers

# 1. Install dependencies
pip install verifiers
uv tool install prime

# 2. Login to Prime
prime login

# 3. Install test environment
prime env install will/wordle

# 4. Test the adapter
python -c "
import asyncio
from environments.verifiers_server import VerifiersEnv, VfEnvConfig
from atroposlib.envs.base import APIServerConfig

config = VfEnvConfig(vf_env_name='wordle')
env = VerifiersEnv(config, [APIServerConfig(model_name='test')])
asyncio.run(env.setup())
print(f'Train: {len(env.train)}, Eval: {len(env.test)}')
"
# Expected: Train: 2000, Eval: 20