voidism/DoLa
A decoding strategy for large language models that reduces hallucinations by contrasting differences in logits across transformer layers.
Not currently ranked — collecting fresh signals.
star history
This repository provides the official implementation of DoLa, a decoding method that improves factuality in LLMs without requiring additional fine-tuning or external knowledge retrieval. The approach contrasts logits between early and later transformer layers to identify and suppress hallucinated tokens. The code integrates with Hugging Face Transformers and includes evaluation scripts for benchmarking factuality improvements.
Frequently asked
- What is voidism/DoLa?
- A decoding strategy for large language models that reduces hallucinations by contrasting differences in logits across transformer layers.
- Is DoLa open source?
- Yes — voidism/DoLa is an open-source project tracked on heatdrop.
- What language is DoLa written in?
- voidism/DoLa is primarily written in Python.
- How popular is DoLa?
- voidism/DoLa has 557 stars on GitHub.
- Where can I find DoLa?
- voidism/DoLa is on GitHub at https://github.com/voidism/DoLa.