openenv-halluguard 2.0.1

Train and evaluate LLMs to avoid hallucinations - 5-line API for RL training with 1M+ examples across 38 datasets

openenv-halluguard

Train and evaluate LLMs to avoid hallucinations.

pip install openenv-halluguard

Quick Start (5 lines)

from hallucination_guard_env import HallucinationEnv

env = HallucinationEnv()
obs = env.reset()
result = env.step(answer="your answer", confidence=0.8)
print(f"Reward: {result.reward}, Hallucinated: {result.is_hallucination}")

What It Does

HallucinationGuard-Env is an OpenEnv RL environment that trains AI models to:

• Answer only from verified context — no fabrication
• Cite real sources — quote verification
• Calibrate confidence — don't be overconfident when wrong

Evaluation Example

from hallucination_guard_env import HallucinationEnv

def my_model(question, context):
    # Call your LLM API here
    # Return answer based ONLY on context
    return "answer from context"

env = HallucinationEnv()
obs = env.reset()
action = my_model(obs.question, obs.context)
result = env.step(answer=action, confidence=0.8)

print(f"Hallucinated: {result.is_hallucination}")
print(f"Reward: {result.reward}")

HTTP API (HuggingFace Space)

For full deployment, use the HuggingFace Space:

import requests

BASE = "https://samsankar-hallucination-guard-env.hf.space"

# Start episode
obs = requests.post(f"{BASE}/reset").json()

# Submit answer
result = requests.post(f"{BASE}/step",
                       json={"answer": "your answer"}).json()

# View leaderboard
lb = requests.get(f"{BASE}/leaderboard").json()

Features

• 38 datasets — SQuAD, HaluEval, TruthfulQA, HotpotQA, MedQA, and more
• Research-grade grader — ROUGE + BERTScore + NLI-DeBERTa
• 3 task difficulties — Beginner → Intermediate → Advanced
• 8 hallucination types — Fabricated facts, false citations, overconfident wrong, etc.

Links

• HuggingFace Space: https://huggingface.co/spaces/SamSankar/hallucination-guard-env
• Interactive Docs: https://samsankar-hallucination-guard-env.hf.space/docs
• Leaderboard: https://samsankar-hallucination-guard-env.hf.space/leaderboard

License

MIT