New RTLCurator method improves LLM training data for RTL code generation

Researchers have developed RTLCurator, a novel method for curating datasets used to train large language models for Register-Transfer Level (RTL) code generation. Existing datasets often suffer from a lack of correctness, with a significant portion of generated code failing functional tests. RTLCurator addresses this by learning a compatibility prior that considers behavioral aspects beyond simple simulation passes, using a small set of validated pairs to calibrate the process. This approach balances alignment, representation coverage, and structural richness, leading to improved model performance even when retaining only 80% of the corpus, outperforming random selection or basic simulation filtering. AI

IMPACT Enhances LLM training data quality for specialized domains like hardware design, potentially improving code generation accuracy.

RANK_REASON The cluster describes a new research paper detailing a novel method for data curation for LLM training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• Codevigo
• DagsHub
• Gotit.pub
• Hugging Face
• RTL
• RTLCoder
• RTLCurator
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →