license: other
task_categories:
- text-classification
tags: - safety
- rl
- alignment
- refusal
- xstest
pretty_name: C-Guard
C-Guard: A Constitution-Grid Instrument for Data-Efficient RL Alignment
Released data and constitution for the paper "A Constitution-Grid Instrument for Data-Efficient RL Alignment (C-Guard)" (COLM 2026 Efficient Reasoning Workshop).
Guards run inline on every LLM turn, so the job is high-volume, short-prompt, and latency-bound. C-Guard asks: given a fixed 4B base model and a read-only XSTest eval, can targeted synthetic data + GRPO shrink over-refusal without opening disguise holes? The constitution is written as one policy per harm topic, crossed with the ways a user can ask; C-LIM, a per-cell learnability score computed on unseen rows, reads the resulting grid and routes each cell to a move — prune a mastered cell, densify a still-learning one, amend a cell whose rule is wrong, or expand the board with a new topic. Every generation feeds RL training with GRPO, and every gain is measured on the trained model.
Contents
| Path | Contents |
|---|---|
data/xstest_prompts.csv |
XSTest eval prompts (450 rows: id, prompt, type, label, focus, note) used as the read-only over-refusal scoreboard |
data/xstest_explorer-v3.html |
Interactive explorer for the XSTest eval results |
constitutions/ |
The constitution: one charter per harm topic (topics/), boundary twin-generation recipes (mechanisms/), and chapters; see constitutions/README.md for the two-axis structure |
License
Released for research use.