Skip to content

Commit dbecb02

Browse files
committed
Fixed image ratios
1 parent 6946108 commit dbecb02

File tree

4 files changed

+4
-4
lines changed

4 files changed

+4
-4
lines changed

AP_sampling/README.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
# Explaining and Improving Contrastive Decoding by Extrapolating the Probabilities of a Huge and Hypothetical LM
22

3-
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/master/AP_sampling/imgs/APD_first_figure.png?raw=true" width="1586" height="1402"></p>
3+
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/master/AP_sampling/imgs/APD_first_figure.png?raw=true" width="540" height="477"></p>
44

55
## Introduction
66

README.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -4,7 +4,7 @@
44

55
Assuming you have a series of LLMs with different sizes that are trained on the same data and you want to increase the factuality and diversity of the text sampled from your largest LLM. Then, consider to use our proposed REAL sampling and/or APD sampling. In FactualityPrompt, we show that APD + REAL sampling outperforms 13 state-of-the-art sampling methods. Our baselines include typical ([Meister et al., 2022](https://arxiv.org/abs/2202.00666)), eta ([Hewitt et al., 2022](https://arxiv.org/pdf/2210.15191)), EDT ([Zhang et al., 2024](https://arxiv.org/abs/2403.14541)), adaptive ([Zhu et al., 2024](https://arxiv.org/abs/2402.18223)), microstat ([Basu et al., 2021](https://arxiv.org/abs/2007.14966)), EAD w/o ELI ([Arora et al., 2023](https://arxiv.org/abs/2302.06784)) factual ([Lee et al., 2022](https://arxiv.org/abs/2206.04624)) top-p ([Holtzman et al., 2020](https://arxiv.org/pdf/1904.09751)), top-k ([Fan et al., 2018](https://arxiv.org/pdf/1805.04833)), and temperature sampling; contrastive search ([Su and Collier, 2022](https://arxiv.org/pdf/2210.14140)) , contrastive decoding (CD) ([Li et al., 2022](https://arxiv.org/pdf/2210.15097)), and DoLa ([Chuang et al., 2023](https://arxiv.org/pdf/2309.03883)). We show that APD + REAL sampling makes Pythia 6.9B simultaneously achieve the factuality of greedy sampling and diversity of top-p with p=0.5.
66

7-
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/main/AP_sampling/imgs/Results.png?raw=true" width="540" height="300"></p>
7+
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/main/AP_sampling/imgs/Results.png?raw=true" width="540" height="195"></p>
88

99
## Usage
1010

REAL_sampling/README.md

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,6 +1,6 @@
11
# REAL Sampling: Boosting Factuality and Diversity of Open-Ended Generation by Extrapolating the Entropy of an Infinitely Large LM
22

3-
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/master/REAL_sampling/imgs/REAL_second_figure.png?raw=true" width="2890" height="622"></p>
3+
<p align="center"><img src="https://github.com/amazon-science/llm-asymptotic-decoding/blob/master/REAL_sampling/imgs/REAL_second_figure.png?raw=true" width="540" height="116"></p>
44

55
## Introduction
66

REAL_sampling/src/train_entropy_prediction_model.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -54,7 +54,7 @@
5454
from transformers.utils import check_min_version, send_example_telemetry
5555
from transformers.utils.versions import require_version
5656
from model import GPTNeoXForEntropyClassification, GPTNeoXForEXPEntropyClassification, GPTNeoXForScaledEntropyClassification
57-
from model import OPTForEntropyClassification, GPT2ForEntropyClassification
57+
from model import OPTForEntropyClassification
5858

5959
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
6060
#check_min_version("4.32.0.dev0")

0 commit comments

Comments
 (0)