license: apache-2.0
task_categories:

• text-generation
  language:
• en
  tags:
• abliteration
• conditioning
• human-written
  size_categories:
• 10K<n<100K
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
  • split: test
    path: data/test-*

harmless-prompts-en-train

English, human-written harmless prompts pooled from OASST2, OASST1 and Aya, the harmless half of an abliteration conditioning pair.

What it is for

The repo holds harmless prompts in ZStandard parquet with a train and a test split, and two columns: text and source.
Abliteration locates a refusal direction from the difference between activations on harmless and harmful prompts, and this is the harmless side.
It replaces Heretic's default harmless set, mlabonne/harmless_alpaca, which derives from Stanford Alpaca under the non-commercial CC-BY-NC-4.0 licence.
Every source here is Apache-2.0 and human-written, not sampled from a model. The source column keeps the origin of every row.
The train split locates the refusal direction and the test split measures refusals afterwards.
The harmful half is chibifire/harmful-prompts-bench, and the code lives in weftspun/request-for-discussion.

Provenance

• OpenAssistant/oasst2, Apache-2.0: English prompter turns.
• OpenAssistant/oasst1, Apache-2.0, revision fdf72ae0827c1cda404aff25b6603abec9e3399b: rows not already in OASST2.
• CohereForAI/aya_dataset, Apache-2.0: English inputs, human-annotated.
• Filters: role == "prompter", lang == "en", text length from 12 to 1000 characters, not deleted, not review_result == false, deduplicated on exact text across all sources.
• Rows are sorted lexicographically, so the order is deterministic and there is no seed.
• A tenth of the rows is held out for test. Assignment is blake2b(text) % 10 == 0, which stays stable when rows are added.
• Citation metadata is in CITATION.cff.

Use

heretic --model <model> \
  --good-prompts.dataset chibifire/harmless-prompts-en-train \
  --good-prompts.column text \
  --bad-prompts.dataset chibifire/harmful-prompts-bench \
  --bad-prompts.column text

from datasets import load_dataset
ds = load_dataset("chibifire/harmless-prompts-en-train")

Licence

Apache-2.0, the licence of all three sources.