HH_length_biased_15k is a 15,000-sample subset of Anthropic/hh-rlhf, created for the paper 'Understanding impacts of human feedback via influence functions'. Taywon Min authored this dataset, which was last updated on December 5, 2024. It contains 976 samples where responses were intentionally flipped to be lengthy.
Use Cases
- Analyzing the impact of response length bias on RLHF training outcomes based on the described flipped samples.
- Studying influence functions in human feedback datasets based on the paper's stated research focus.
- Benchmarking language model training with artificially induced length preferences based on the dataset's construction.
Strengths
- Dataset size is explicitly stated as 15,000 samples.
- Contains 976 specifically manipulated samples to induce length bias.
- Derived from the established Anthropic/hh-rlhf dataset.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Last updated 2024-12-05 01:05:11; freshness should be verified.
Provenance
- Source
- Anthropic/hh-rlhf
- Collection Method
- A random subset with 976 samples flipped towards lengthy responses.
- Freshness
- 2024-12-05 01:05:11