Safety & AlignmentSaturated
Trust Score:
25

TruthfulQA

817 adversarial questions probing whether models repeat common human misconceptions — famous for finding that bigger models were often less truthful.

Launched: —Refresh: static
Status Assessment (saturated):

By 2023 post-RLHF models exceeded the 94% human truthfulness baseline — but largely because the exact misconceptions TruthfulQA targets are what alignment tuning trains against, and the dataset itself leaked into alignment mixes. Saturation here reflects targeted contamination, not broadly truthful models. saturated_date (2023-06) marks the ChatGPT/GPT-4 era when the inverse-scaling trend reversed.

Performance Timeline

Longitudinal progression of model scores against human baselines.

Performance & Historical Trajectory

Empirical score progression across model release dates and evaluation rounds.

Independent Vendor-Reported Human Baseline (94.0%)
ModelScoreDateSource TypeProvenance
GPT-3 175B58%2021-09-08independentSource ↗

Human Baseline & Difficulty Horizon

Calibrated human reference points, specialist benchmarks, and ceiling thresholds.
Measured Human Score94.0%Domain Expert Baseline
Baseline Protocol & Interpretation

Human crowdworker truthfulness rate on questions designed to elicit common misconceptions.

Metric & Scoring Methodology

Verification protocols, aggregation formulas, and specialized metric variants.
Primary Metric:MC2 (normalized probability mass on true answers, %)
Scoring Engine:log-likelihood

Dataset & Compute Cost

Evaluation volume, public availability, API pricing, and local hardware requirements.
Total Dataset Size817Annotated evaluation items
Public Test SetPublicOpenly mirrored on repositories
Access GatingOpen AccessUnrestricted download
Evaluation LicenseOpen AccessDataset usage and redistribution terms
Frontier API Compute Cost:

$5 – $20 USD for full benchmark evaluation run on frontier APIs.

Recommended Local GPU Setup:

1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang

Official Dataset & Benchmark Files:Download / View Dataset Repository ↗

How to Run & Reproduce

Standardized evaluation protocols, CLI commands, and reproducible runner templates.
Prompt Regimezero-shot
Reasoning Modedirect
Sampling Temp0
Pass@k Budgetk = 1
Tools & SandboxPure Text
Scoring Verifierlog-likelihood
Option AEleutherAI LM-Evaluation-Harness (Open-Weight Models)
lm_eval --model hf --model_args pretrained=<model_path> --tasks truthfulqa --batch_size auto
Option BOpenCompass Evaluation Framework
opencompass --datasets truthfulqa --models <model_config>
Python APIDeterministic Inference Loop Snippet
# Standard API Evaluation Loop
from openai import OpenAI

client = OpenAI()
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": prompt}],
    temperature=0.0,
)
Standardized Reporting Requirement:

When publishing results for TruthfulQA, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.

Contamination & Memorization Analysis

Audit of pretraining exposure risks, memorization vectors, and refresh policies.
Overall Contamination Risk:HIGH
Refresh Cadence:

Static fixed snapshot

Test Set Exposure:

Public on web / HuggingFace