PaulR11/training-runs-concerningness
GRPO Training Runs — Concerningness Scalar Reward (8-Target) Part of the data release for "Training Alignment Auditors via Reinforcement Learning" (ICLR 2026). Scalar-reward GRPO using the 38-metric Petri judge to compute a concerningness-based reward R = gate(realism) × ((c-3)/7)² with penalties for auditor_failure, stuck_in_loops, and eval_awareness. Uses 128-trajectory batches (vs 64 for other runs), so compute-matched epoch is halved.… See the full description on the dataset page: https://huggingface.co/datasets/PaulR11/training-runs-concerningness.
| Author / organization | PaulR11 |
|---|---|
| License (as declared) | other |
| Downloads | 328 (as reported by Hugging Face) |
| Likes | 0 |
| Created | 24 April 2026 |
| Last updated at source | 24 April 2026 |
Popularity, last 3 days
328 now. Counts as reported by the source, snapshotted daily by Silytics.
Recent changes
- 16 Oct 2026 Description: GRPO Training Runs — Concerningness Scalar Reward (8-Target) Part of the data release for "Training Alignment Auditors via Reinforcement Learning" (ICLR 2026). Scalar-reward GRPO using the 38-metric Petri judge to compute a concerningness-b → GRPO Training Runs — Concerningness Scalar Reward (8-Target) Part of the data release for "Training Alignment Auditors via Reinforcement Learning" (ICLR 2026). Scalar-reward GRPO using the 38-metric Petri judge to compute a concerningness-b
Changes detected by Silytics between refreshes of the source.
More from PaulR11
Silytics shows public metadata from Hugging Face and links to the original. Information may be incomplete or out of date. Last refreshed 18 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.