---
pretty_name: RuleFollower Unified Sentence G/F Automatic Annotations
language: en
configs:
- config_name: sentence_labels
  default: true
  data_files:
  - split: train
    path: sentence_labels/*.jsonl.gz
- config_name: cots
  data_files:
  - split: train
    path: cots/*.jsonl.gz
---

# RuleFollower unified G/F automatic annotations

This is a **training-only, model-annotated subset**. It contains 1,983 source IDs from 16 datasets, four CoT generators per source (7,932 CoTs), and 251,870 sentence-label rows. The original 2,000-source sample used dataset-size-proportional quotas across 55,965 source rows.

`cots/` contains the task, classification input, complete CoT, and canonical sentence IDs. `sentence_labels/` contains sentence-level labels and evidence; join the two by `cot_id` and use `sid` within a CoT. `source_manifest.csv` lists the included source IDs and input-text hashes. `metadata.json` gives exact counts by generator.

`G=1` means the Luna teacher confirmed grounding in the classification input; `G=0` means it did not confirm grounding. `F=1` means the frozen N1 pipeline confirmed a fallacy; `F=0` means it did not confirm one after successfully processing the CoT. Neither zero is a human-verified opposite class. `null` is a technical abstention: use the separate `training_masks.G` and `training_masks.F` fields when computing losses. 

Teacher: GPT-5.6-Luna (medium), direct-v3 G and frozen N1-r12 F. Exact input-text matches to the locked 160-source human pool were excluded before packaging (17 source IDs across each generator). Human test labels are not included. Exact duplicate texts may still occur within this training subset; group by input-text hash when splitting or sampling. 

The classification inputs derive from [RuleFollower's parsed datasets](https://huggingface.co/datasets/RuleFollower/rulefollower_parsed). Review source-dataset redistribution terms before making this subset public.