indirected / llm-earlyexit

An Early Exit Analysis for LLMs. Part of the NLP course.

LLM Early Exit

An Early Exit strategy Analysis for LLMs

Usage

You can install the required dependencies with the following command:

conda env create -f environment.yml

Then activate the environment:

conda activate torch-main

Early Layer Visualization

run and follow the cells in the vis.ipynb file.

an Example visualization of the probability distribution of a wrong answer is:

• Question: How many whiskers does a human have?
• Correct Answer None
• Model Generation: Two

Entropy Visualization

The entropy of layers based on correct and wrong answers on the TruthfulQA dataset shows potential bias (low entropy) in early layers.

The code can be found in the Entropy.ipynb file.

MLP Fact Detection

We train simple MLPs on the inner layer logits on the true-false dataset in the 'classifier.ipynb` file to try to detect unfactual generations.

Here are some results:

Experiment	Train Loss	Train Acc	Test Acc
LM Next Token	-	0.0	0.0
LM Token Prob.		0.50	0.53
MLP tap at layer 16	0.6228	0.56	0.65
MLP tap at layer 20	0.6283	0.56	0.63
MLP tap at layer 24	0.6263	0.55	0.61
MLP majority	-	0.56	0.63

A Potential Solution

We use the DoLa decoding Method to generate new answers. Although the new generations seem better to humans. These automated metrics below seem to not agree. More information can be found at the Original DoLa Repository.

	BLEU	ROUGE	BERTScore
Vanilla Decoding	0.244445	0.409804	0.900487
Dynamic-DOLA	0.16386	0.347001	0.891913