Toxic-DPO: Preference Data for Model Unalignment

Unalignment Toxic Dpo V0.2 Zh Cn is a multilingual dataset intended to illustrate the use of Direct Preference Optimization (DPO) for model unalignment. The dataset was created by tastypear and last updated on 2024-01-31. Its description states it contains highly toxic or harmful examples.

Use Cases

• Training preference models for content moderation based on the described toxic examples.
• Studying the effects of DPO on model behavior using the described harmful preference pairs.
• Benchmarking model alignment techniques against adversarial preference data.
• Creating multilingual adversarial test sets for safety evaluation based on the Chinese-English content.

Strengths

• Dataset is explicitly designed for a specific machine learning technique (DPO).
• Provides a Chinese-English parallel version, as stated in the description.
• Last update timestamp is precisely recorded as 2024-01-31 13:57:28.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

huggingface user tastypear

Collection Method

Likely generated or curated for DPO experiments; Chinese version created via model paraphrasing.

Freshness

Last updated 2024-01-31 13:57:28; freshness should be verified.