README.md

---
license: cc-by-sa-3.0
task_categories:
- text-classification
language:
- en
tags:
- hate-speech
- misinformation
- check-worthiness
- llm-in-the-loop
- llm-as-annotator
pretty_name: WSF-ARG+


configs:
- config_name: per_claim
  data_files:
  - split: train
    path: wsf_arg_plus_per_claim.csv

- config_name: per_claim_all_llms
  data_files:
  - split: train
    path: wsf_arg_plus_per_claim_all_llms.csv

- config_name: per_claim_gold_disagg
  data_files:
  - split: train
    path: wsf_arg_plus_per_claim_gold_disagg.csv

- config_name: per_claim_platinum_disagg
  data_files:
  - split: train
    path: wsf_arg_plus_per_claim_platinum_disagg.csv

- config_name: per_message
  data_files:
  - split: train
    path: wsf_arg_plus_per_message.csv

- config_name: per_message_all_llms
  data_files:
  - split: train
    path: wsf_arg_plus_per_message_all_llms.csv

- config_name: per_message_gold_disagg
  data_files:
  - split: train
    path: wsf_arg_plus_per_message_gold_disagg.csv

- config_name: per_message_platinum_disagg
  data_files:
  - split: train
    path: wsf_arg_plus_per_message_platinum_disagg.csv
---

# WSF-ARG+

This dataset card describes the **WSF-ARG+** dataset presented in the paper *"When Hate Meets Facts: LLMs-in-the-Loop for Checkworthiness Detection in Hate Speech,"* accepted to the main conference of EMNLP 2026.

We organized the data into two tabular formats:

**Message-level format**: In this format, each row corresponds to an entire message, which can be either hate speech or non-hate speech. Each message is divided into one or more claims. If the message is argumentative, its premises and conclusion are explicitly annotated; otherwise, it is stored as one or more claims. Consequently, each row represents a full message, with columns corresponding to the individual claims. Additional information includes:
- The overall hatefulness of the message
- Whether the message is argumentative
- Annotations per claim (hatefulness and check-worthiness labels)
- Existing annotations from WSF-ARG

The file `wsf_arg_plus_per_message.csv` contains the aggreggated labels through majority voting. It has both gold (obtained through LLM-in-the-loop) and platinum (obtained through full human annotation) check-worthiness annotations. `wsf_arg_plus_per_message_gold_disagg.csv` and `wsf_arg_plus_per_message_platinum_disagg.csv` contain the annotations disaggregated given by all annotators for both gold and platinum. We also indicate which claims required to be judged. We also released `wsf_arg_plus_per_message_all_llms.csv` that contains all the predictions carried out in the check-worthiness detection task for all configurations (per LLM and prompt). We release the three runs per configuration and the majority voting label. Consider that we have 12 LLMs tested, and 2 prompt strategies being in total 12x2 = 24 configurations. Each configuration is run three times meaning a total of 24x3 = 96 runs. Therefore, the .csv file contains 96 columns with each of these runs and 12 other columns for the majority voting label.

**Claim-level format**: This format transforms the message-level table into one where each row represents a single claim. Each claim may originate from a full message. For each claim, we record:
- The message it comes from
- Whether the claim itself is hateful
- Whether it comes from a hate speech or non-hate speech message
- Check-worthiness annotations provided by the LLM-in-the-loop
- Annotations from all human annotators
- Annotations from each model included in our experimental study

Similarly to the message-level format, we have:

`wsf_arg_plus_per_claim.csv` as the general dataset organized in claim-level format.
`wsf_arg_plus_per_claim_gold_disagg.csv` and `wsf_arg_plus_per_claim_platinum_disagg.csv` for the disaggregated gold and platinum labels.
`wsf_arg_plus_per_claim_all_llms.csv` that contains the disaggregated and aggregated annotations carried out by all our 24 configuration (12 LLMs and 2 prompting strategies).