Response ratio predicts instruction-tuning utility better than common proxies
Original title: "High-Quality" Data for Early-Stage Instruction Tuning: An Empirical Study of Data Quality Metrics
Under a fixed finetuning budget, answer-dense examples beat common quality proxies on perplexity and LLM-as-judge scores.
A paper treats data quality as an operational property of instruction-tuning examples. Simple example-level metrics from data selection, text quality, and instruction tuning are scored, then examples are split into low, middle, and high bins of roughly equal token mass and finetuned separately. Validation perplexity and LLM-as-judge scores are compared against cumulative training tokens. The authors report that response ratio, capturing answer-dense supervision, is a stronger and more portable predictor of training utility than commonly used alternative proxies.
Unconfirmed: The ranking of metrics is the authors' result from this controlled comparison, not an independent replication.
- Bins are formed so each has roughly equal token mass.
- Models trained on each bin are compared by validation perplexity and LLM-as-judge scores.
- Utility is measured as a function of cumulative training tokens.
- The comparison uses a fixed finetuning budget.
From the original article
1 min read at arxiv.org"Data quality" is widely invoked in instruction tuning, but the term is often underspecified. We study it as an operational property by asking which simple example-level metrics, drawn from prior work on data selection, text quality, and instruction tuning, best predict training utility under a fixed finetuning budget.
By Vladimir Necula · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.