gitwtfhub

wtf is truthrl?

facebookresearch/truthrl — explained in plain English

Analysis updated 2026-07-17 · repo last pushed 2025-12-10

★ 17PythonAudience · researcherComplexity · 5/5QuietSetup · hard

TL;DR

TruthRL trains language models to say "I don't know" instead of making up answers, using reinforcement learning that rewards honesty and penalizes hallucinations.

Mindmap

mindmap
  root((repo))
    What it does
      Reduces hallucinations
      Rewards honest abstention
      Uses RL training
    Tech stack
      Python
      Reinforcement learning
      LLM judge
    Use cases
      Trustworthy QA models
      Uncertainty-aware chatbots
      RL reward research
    Audience
      ML researchers
      RL practitioners
    Requirements
      8 high-end GPUs
      Judge LLM access

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Train a language model to abstain from answering when it's likely to be wrong instead of hallucinating.

REASON 2

Set up an LLM-as-judge reward loop that penalizes made-up answers and rewards correct ones.

REASON 3

Research reinforcement-learning techniques for making QA systems more trustworthy.

What's in the stack?

PythonReinforcement Learning

How it stacks up

facebookresearch/truthrl0petru/sentimoalingalingling/akasha-wechat
Stars171717
LanguagePythonPythonPython
Last pushed2025-12-10——
MaintenanceQuiet——
Setup difficultyhardmoderatehard
Complexity5/53/54/5
Audienceresearcherdeveloperdeveloper

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1day+

Requires 8 high-end GPUs plus a separate LLM to act as judge/verifier.

Yoink these prompts

Prompt 1
Explain how TruthRL's reward function treats correct answers, hallucinations, and abstentions differently, and help me adapt it for my own model.
Prompt 2
Walk me through setting up TruthRL's RL training pipeline, including the GPU and judge-LLM requirements.
Prompt 3
Show me how to use an LLM as a judge to score my model's answers as correct, hallucinated, or abstained.
Prompt 4
Summarize the core idea behind TruthRL's approach to reducing hallucinations in LLMs.

Frequently asked questions

wtf is truthrl?

TruthRL trains language models to say "I don't know" instead of making up answers, using reinforcement learning that rewards honesty and penalizes hallucinations.

What language is truthrl written in?

Mainly Python. The stack also includes Python, Reinforcement Learning.

Is truthrl actively maintained?

Quiet — no commits in 6-12 months (last push 2025-12-10).

How hard is truthrl to set up?

Setup difficulty is rated hard, with roughly 1day+ to a first successful run.

Who is truthrl for?

Mainly researcher.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.

Other repos in this lane