Get the app
misc 0 0 0

ceselder/loracle-syntactic-triggers-v1

published by @ceselder

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
179
Last 30 days
22
Likes
-
Size
70.6 MB
Created on HF
2026-04-30
Age
5 months ago

Description

Syntactic-Trigger Backdoors (v1)

2616 distinct (trigger, behavior) backdoor specs designed to expand the IA

backdoor corpus beyond DiT's single SEP-prefix trigger style. Generated via

archetype-constrained Sonnet 4.6 prompting (176 archetypes × 15 specs each)

to guarantee structural diversity across 14 trigger axes.

Each LoRA was trained on Qwen3-14B (rank 4, alpha 8, 30 pairs × 2 epochs).

Files

specs.parquet — 2616 rows: spec_id, archetype info… See the full description on the dataset page: https://huggingface.co/datasets/ceselder/loracle-syntactic-triggers-v1.

Tags

task_categories:text-generationlanguage:enlicense:mitsize_categories:100K<n<1Mformat:parquetmodality:textlibrary:datasetslibrary:pandaslibrary:polarslibrary:mlcroissantregion:usbackdoorlorasafetymechinterp

README current version from Hugging Face

license: mit
task_categories:

  • text-generation
    language:
  • en
    tags:
  • backdoor
  • lora
  • safety
  • mechinterp
    size_categories:
  • 1K<n<10K

Syntactic-Trigger Backdoors (v1)

2616 distinct (trigger, behavior) backdoor specs designed to expand the IA
backdoor corpus beyond DiT's single SEP-prefix trigger style. Generated via
archetype-constrained Sonnet 4.6 prompting (176 archetypes × 15 specs each)
to guarantee structural diversity across 14 trigger axes.

Each LoRA was trained on Qwen3-14B (rank 4, alpha 8, 30 pairs × 2 epochs).

Files

  • specs.parquet — 2616 rows: spec_id, archetype info, trigger_description,
    behavior_description, examples, train/val split (90/10).
  • training_pairs.parquet — 152,278 rows of (trigger_prompt, trigger_response,
    plain_prompt, base_response) used to train each LoRA. spec_id joins to specs.

Axes (14 total)

LEXICAL (284), PHRASAL (284), LANGUAGE/SCRIPT (282), STRUCTURAL (281),
ADVERSARIAL/INJECTION (250), LINGUISTIC (194), CODE/MARKUP (180),
TEMPORAL/SESSION (165), POSITIONAL (150), NUMERIC/FORMAT (135),
UNICODE-INVISIBLE (119), SEMANTIC-ROLE (114), META-CONTEXT (90),
TYPOGRAPHIC (88).

Generation pipeline

  • Spec gen: Sonnet 4.6 (anthropic/claude-sonnet-4.6), archetype-driven
  • Pair gen: Gemini 3.1 Flash Lite Preview (google/gemini-3.1-flash-lite-preview)
  • Trainer: paper_ablations/syntactic_triggers/ in lora-oracles repo

Companion artifacts

  • LoRAs themselves (2615 PEFT adapters): not yet on HF
  • Direction tokens (svd_fixed_k16_mag7_rankfirst): not yet on HF

README history 2 revisions

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.

  1. 2026-04-30Initial upload: 2616 syntactic-trigger backdoor specs + 152K training pairs23e9c251.5 KB
    Loading...
  2. 2026-04-30initial commit8b167910 B
    Loading...

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.

Related datasets same stage · misc

nvidia/Aegis-AI-Content-Safety-Dataset-2.0
99.3K dl 114 likes 27.1 MB
🛡️ Nemotron Content Safety Dataset V2 The Nemotron Content Safety Dataset V2, formerly known as Aegis AI Content Safety Data...
allenai/tulu-3-sft-personas-instruction-following
88.4K dl 70 likes 37.4 MB
Dataset Descriptions This dataset contains 29980 examples and is synthetically created to enhance model's capabilities to follo...
meta-llama/Llama-3.1-8B-Instruct-evals
73.2K dl 36 likes 232.9 MB
Dataset Card for Llama-3.1-8B-Instruct Evaluation Result Details This dataset contains the Meta evaluation result details for...
Manusagents/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-5-Mythos-5-Qwen-3.7-Max-and-more-Distillation-Dataset
58.3K dl 226 likes 10.8 GB
📖 The Open Distillation Codex 🌌 The Ultimate Open-Source Distillation Dataset — No Skip, Full, with Attack & Defe...
nvidia/Aegis-AI-Content-Safety-Dataset-1.0
36.1K dl 63 likes 5.2 MB
🛡️ Nemotron Content Safety Dataset V1 Nemotron Content Safety Dataset V1, formerly known as Aegis AI Content Safety Dataset, i...
innodatalabs/rt-gsm8k-gaia
35.9K dl 15 KB
rt-gsm8k dataset Red teaming gsm8k-gaia dataset. Generated from the "GSM8K" dataset and "GAIA" dataset. Dataset Str...
HuggingFaceH4/hhh_alignment
35.3K dl 27 likes 266 KB
This task evaluates language models on alignment, broken down into categories of helpfulness, honesty/accuracy, harmlessness, and ...
ethz-spylab/rlhf_trojan_dataset
26.2K dl 6 likes 29.9 MB
Dataset for the competition This is the official dataset for the competition "Find the Trojan: Universal Backdoor Detection in ...

Read further

Community best practices (architecture)
4 rules for MoE, hybrid SSM, FP8 requantization, and Gemma quirks.
The six workflow stages
What misc is doing more work than its name suggests.
Beyond English: multilingual corpora
The four language communities and their workarounds.
Non-English abliteration communities
Deep dive: huihui-ai, layer-window, slop-framing, CJK tokenizer fixes.

Metadata

License
mit

Repository

FILE .gitattributes FILE README.md FILE specs.parquet FILE training_pairs.parquet

File contents live on Hugging Face. Click a file or folder to open it there.

Download the dataset

0

Copy any of these snippets into your notebook or terminal. All three fetch directly from Hugging Face using your own credentials.

Recommended: datasets library (Python) 
from datasets import load_dataset
ds = load_dataset("ceselder/loracle-syntactic-triggers-v1")
Raw snapshot (Python) 
from huggingface_hub import snapshot_download
snapshot_download(repo_id="ceselder/loracle-syntactic-triggers-v1", repo_type="dataset")
git clone (requires git-lfs) 
git clone https://huggingface.co/datasets/ceselder/loracle-syntactic-triggers-v1
Browse repository on Hugging Face ->
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.

Get the app Browse the catalog
Open in Abliteration