voidism / DoLa

Official implementation for the paper "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models"

https://arxiv.org/abs/2309.03883

voidism/DoLa Issues

concerns about your experiments of TruthfulQA-MC
Closed 8 months ago2
The way to calculate the log_probs
Updated 4 months ago
data for gpt-judge training
Closed 4 months ago1
Support for LLaMA-2
Closed 8 months ago15
Typo? mode = "early_exit_contrastive" -> mode = "dola-static"
Closed 5 months ago1
Some questions about the idea of the paper
Closed 5 months ago1
How to install the environment if I don't have root permission?
Updated 5 months ago
Query regarding JS Divergence mean over batches.
Closed 6 months ago2
Replacing .sum() with .mean() in dola.py?
Updated 6 months ago2
What tool do you use to get the token prediction of each layer of large language models for Figure 2?
Closed 6 months ago3
dola可以适配xverse模型吗？
Closed 7 months ago1
Isn't the range of the JS divergence [0,1]? Why are there values greater than 1 in Figure 2?
Closed 8 months ago1
questions about code
Closed 8 months ago1
Should apply model.norm layer to hidden_states[early_exit_layer] ?
Updated 9 months ago1
Figure 2 subtoken "_"
Closed a year ago