Get the app
extract 0 0 0

eulogik/suraksha-guard-38k

published by @eulogik

Alignment dataset tracked in the /datasets sub-catalog.

Lifetime downloads
160
Last 30 days
160
Likes
1
Size
15.5 MB
Created on HF
2026-10-09
Age
12 days ago

Description

suraksha-guard-38k: frozen guard training data (English + Hinglish + Hindi)

Guard mix behind SURAKSHA 450M by Eulogik. Template-generated with ground truth by construction: tool-risk labels come from the generated tool JSON, injection labels come from the template. No LLM-judge labeling. Code: github.com/eulogik/suraksha.

Files

guard_train.jsonl: 24,252 unique rows (deduplicated from 38,000 generated).

guard_test.jsonl: 3,058 unique rows, disjoint templates… See the full description on the dataset page: https://huggingface.co/datasets/eulogik/suraksha-guard-38k.

Tags

task_categories:text-classificationlanguage:enlanguage:hilicense:apache-2.0size_categories:10K<n<100Kformat:parquetmodality:textlibrary:datasetslibrary:pandaslibrary:polarslibrary:mlcroissantregion:usguardrailsprompt-injectionjailbreakpiimcphinglishsynthetic

README current version from Hugging Face

license: apache-2.0
language:

  • en
  • hi
    tags:
  • guardrails
  • prompt-injection
  • jailbreak
  • pii
  • mcp
  • hinglish
  • synthetic
    task_categories:
  • text-classification

suraksha-guard-38k: frozen guard training data (English + Hinglish + Hindi)

Guard mix behind SURAKSHA 450M by Eulogik. Template-generated with ground truth by construction: tool-risk labels come from the generated tool JSON, injection labels come from the template. No LLM-judge labeling. Code: github.com/eulogik/suraksha.

Files

  • guard_train.jsonl: 24,252 unique rows (deduplicated from 38,000 generated).
  • guard_test.jsonl: 3,058 unique rows, disjoint templates, unseen tools, unseen exfil domains.
  • SHA256SUMS, LICENSES.json: freeze bundle pinned before training.

Row shape: {state, questions, expected, labels, split}. Prompt rows carry prompt_injection (bool) plus severity (0-2). Tool rows carry tool_risk (safe/write/privileged/exfiltrate) plus severity.

Coverage

English plus Roman Hinglish plus Devanagari plus code-switch. PII is synthetic only (faker-style PAN/Aadhaar/phone/keys, plain plus split plus redacted plus base64). No verbatim real PII. Test tools (shell_exec, db_query, send_email, upload_file) never appear in train. Test domains never appear in train.

License

Apache-2.0 (ours, synthetic). Banking77 rehearsal rows used in training are PolyAI CC-BY-4.0 and live outside this dataset.

Keywords: guardrails dataset, prompt injection data, hinglish dataset, mcp tool risk data, synthetic security data.

README history 8 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. 2026-10-09Upload guard_test.jsonl with huggingface_hubb799d371.6 KB
    Loading...
  2. 2026-10-09Upload guard_train.jsonl with huggingface_hubfe558e11.6 KB
    Loading...
  3. 2026-10-09Upload README.md with huggingface_hubfc1a7d81.6 KB
    Loading...
  4. 2026-10-09Upload LICENSES.json with huggingface_hub3d9366b0 B
    Loading...
  5. 2026-10-09Upload SHA256SUMS with huggingface_hub32e163a0 B
    Loading...
  6. 2026-10-09Upload guard_test.parquet with huggingface_hubfacc1d00 B
    Loading...
  7. 2026-10-09Upload guard_train.parquet with huggingface_hub44901b40 B
    Loading...
  8. 2026-10-09initial commitf2fc38d0 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 · extract

Anthropic/hh-rlhf
2.1M dl 2442 likes 90.4 MB
Dataset Card for HH-RLHF Dataset Summary This repository provides access to two different kinds of data: Human ...
djain95/sae-jailbreaks-cache
468.6K dl 1 likes 30.6 GB
mosaicml/dolly_hhrlhf
309.5K dl 113 likes 23.7 MB
Dataset Card for "dolly_hhrlhf" This dataset is a combination of Databrick's dolly-15k dataset and a filtered subset of Anthrop...
Dahoas/full-hh-rlhf
266.3K dl 104 likes 129.8 MB
Dataset Card for "full-hh-rlhf" Anthropic's HH dataset reformatted into prompt, chosen, rejected samples.
mlabonne/harmful_behaviors
198.2K dl 169 likes 23 KB
mlabonne/harmless_alpaca
190.2K dl 51 likes 1.2 MB
walledai/AdvBench
176.3K dl 123 likes 38 KB
Dataset Card for AdvBench Paper: Universal and Transferable Adversarial Attacks on Aligned Language Models Data: AdvBench Datas...
JailbreakV-28K/JailBreakV-28k
116.4K dl 80 likes 288.8 MB
⛓‍💥 JailBreakV-28K: A Benchmark for Assessing the Robustness of MultiModal Large Language Models against Jailbreak Attacks 🌐 ...

Read further

Community best practices (extraction)
5 rules the field converged on for building the pair.
Building your own extraction pair
When the reference pair is not enough. Sizing, framing, and evaluation.
Datasets that shaped abliteration
The five corpora that did most of the work.
Non-English abliteration communities
How Chinese, Russian, Japanese, and Korean practitioners build extraction pairs.

Metadata

License
apache-2.0

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("eulogik/suraksha-guard-38k")
Raw snapshot (Python) 
from huggingface_hub import snapshot_download
snapshot_download(repo_id="eulogik/suraksha-guard-38k", repo_type="dataset")
git clone (requires git-lfs) 
git clone https://huggingface.co/datasets/eulogik/suraksha-guard-38k
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