News-Truthful
Harbor: pendingThis the dataset for Every Language Counts: Learn and Unlearn in Multilingual LLMs.
- Samples
- N/A
- Released
- 2024
- Scores
- 0
$ npm i -g lettokenburn-cli$ ltb bench pull news-qa
External harness. pip install crfm-helm && helm-run --run-entries newsqa:model=<model> --suite my-suite --max-eval-instances 100 (HELM scenario 'newsqa'; some scenarios need extra args) https://github.com/stanford-crfm/helm
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| OpenAI/o3-2025-01-31-high | 92.5 | N/A | Independent | medium | Jan 30, 2025 | arxiv.org |
| OpenAI/o4-mini-2025-05-01-high | 88.2 | N/A | Independent | medium | Jun 13, 2021 | arxiv.org |
| OpenAI/o1-2024-12-17-high | 81.4 | N/A | Independent | medium | Jun 19, 2022 | arxiv.org |
| DeepSeek R1 | 80.6 | N/A | Independent | medium | Jan 22, 2025 | arxiv.org |
| Anthropic/claude-3-7-sonnet | 74.2 | N/A | Independent | medium | Jun 24, 2024 | www-cdn.anthropic.com |
| xAI/grok-3-1212 | 70.6 | N/A | Independent | medium | Feb 15, 2022 | arxiv.org |
| GPT-4o | 70.2 | N/A | Independent | medium | Oct 3, 2024 | arxiv.org |
| Gemini 2.5 Pro | 68.8 | N/A | Independent | medium | Mar 8, 2024 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
The NewsQA dataset is from the paper: https://arxiv.org/abs/1611.09830
Original repository can be found at: https://github.com/Maluuba/newsqa
Note: The training dataset cannot be directly shared due to copyright issues, and needs to be downloaded by following the instructions in the repo above. These instructions are duplicated here for convenience.
You need to create a login account to download this data.
This should result in the creation of the file (combined-newsqa-data-v1.json) in the repo which is used in this scenario.
NewsQA is a QA dataset containing 12,744 stories, and over 119,633 question-answer pairs. There are 92549 training qa pairs, 5166 qas in the dev set, and 5126 in the test set. Particularly, given the a news article from CNN, the goal is answer questions with answers consisting of spans of text from the corresponding articles. All of the questions and answers are written by crowd sourced human annotators. For more details, see https://arxiv.org/abs/1611.09830.
More concretely, we prompt models using the following format
Passage: <news article> Question: <question> Answer:
Note: Some of the questions do not have an answer in the context so the model needs to answer "No Answer". While this behavior might be tricky to learn in the few-shot setting, we still include these examples in the scenario.
Using an example from the training dataset, we have:
NEW DELHI, India (CNN) -- A high court in northern India on Friday acquitted a wealthy businessman facing the death sentence for the killing of a teen in a case dubbed 'the house of horrors.' Moninder Singh Pandher was sentenced to death by a lower court in February... Question: Who was sentenced to death in February? Answer:
References
['Moninder Singh Pandher']