Full HH RLHF: Anthropic's Human Preference Data Reformatted for Training

Anthropic's HH dataset reformatted into prompt, chosen, and rejected samples by Dahoas. The data was last updated on Hugging Face in February 2023. It provides a structured format for training and evaluating language models using human preferences.

Use Cases

• Training reward models based on human preference data.
• Fine-tuning language models via reinforcement learning using the chosen and rejected samples.
• Benchmarking model alignment techniques against a known human feedback dataset.
• Studying the characteristics of helpful and harmless assistant responses.

Strengths

• Derived from Anthropic's established HH dataset, a known source for human preference data.
• Reformatted specifically for RLHF training workflows with prompt, chosen, and rejected samples.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last updated 2023-02-23 17:29:46; freshness should be verified.

Provenance

Source

Anthropic's HH dataset, reformatted by Dahoas.

Collection Method

Reformatted from the original dataset structure.

Freshness

2023-02-23 17:29:46