Safety benchmarks, over-refusal tests, and guard-model data.
language:enlanguage:hilanguage:zhlanguage:mrlicense:otherregion:ussafetymultilingual+2 more
Lifetime downloads
65
Last 30 days
65
Likes
-
Size
75.1 MB
Description
tiny-aya-safety-drift-repair
Content note. Built for safety research in the Lexsi Labs AI Safety & Alignment Hackathon (Track 2b). This repository contains or enables harmful model behaviour. Do not deploy.
Everything behind the study "We found the refusal circuit. It was the wrong place to repair" on Tiny Aya Global.
data/: parallel five-language (EN, HI, ZH, MR, Hinglish) evaluation prompts (harmful_eval.jsonl from SORRY-Bench-Multilingual, xstest_safe.jsonl… See the full description on the dataset page: https://huggingface.co/datasets/Raj-taware/tiny-aya-safety-drift-repair.
Used by abliterated models
none yet
No abliterated models in our catalog have declared this dataset in their YAML metadata yet. This can mean the dataset is used but not declared, is used outside abliteration workflows, or is new. The link will populate automatically as models are indexed.
Content note. Built for safety research in the Lexsi Labs AI Safety & Alignment Hackathon (Track 2b). This repository contains or enables harmful model behaviour. Do not deploy.
Everything behind the study "We found the refusal circuit. It was the wrong place to repair" on Tiny Aya Global.
data/: parallel five-language (EN, HI, ZH, MR, Hinglish) evaluation prompts (harmful_eval.jsonl from SORRY-Bench-Multilingual, xstest_safe.jsonl, xstest_tune.jsonl), MASSIVE train/test subsets, the harmful fine-tuning mix (harmful_ft.jsonl), the two extra un-align sets, the human judge-check labels (handlabel_gate0.csv) and manifest.json (source revisions, seeds).
runs/<stage>/: per-stage kernel outputs: raw generations (gen_*.jsonl), WildGuard-judged generations (judged_*.jsonl) and result JSON.
results/: aggregate statistics per stage. predictions/: the pre-registered predictions.
RUNLOG.md, DECISIONS.md: what every run returned, and its design written before the run.
MR prompts are machine-translated; Hinglish prompts were written by language models. Source datasets keep their own licences (SORRY-Bench, XSTest, MASSIVE, LLM-LAT/harmful-dataset, BeaverTails).
Every night we snapshot the README of every dataset in the catalog. When the SHA changes we archive the new version and diff it against the last. This is the evolving thought record of the alignment-data field: what the author decided to say about the corpus, and how that framing shifted over time.
2026-11-11Upload folder using huggingface_hub05cbcce1.4 KB
Loading...
2026-11-11Upload folder using huggingface_hub39c00431.4 KB
Loading...
2026-11-11initial commita39c77b0 B
Loading...
Catalog is the map. Apps are the tools.
Run models on your own machine, not in the cloud.
Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.