Koala: A Dialogue Model for Academic Research
Koala:
A Dialogue Model for Academic Research
Introduction
The
accidental leak of the weights associated with Meta AI’s LLM LLaMA has sparked
tremendous innovation in the open-source LLM space. Since the furious leak, we
have seen models like Alpaca, Vicuna, ChatLlama, and several others expand on
the foundations of LLaMA to build innovative conversational agents that match
the capabilities of ChatGPT. One of the latest additions to the list is Koala
(yes, I know, another animal-named model), a chatbot created by Berkeley AI Research
(BAIR) that fine-tunes LLaMA on conversations gathered from the internet.
The
core goal of Koala is to highlight the balance between mega-large closed-source
LLMs and smaller, open-source LLMs. BAIR’s thesis is that smaller models can
achieve performance that matches mega models like ChatGPT with a fraction of
the cost while also improving in areas such as fine-tuning, transparency, and
many others.
Koala
Koala
is a version of LLaMA fine-tuned on dialogue data scraped from the web and
public datasets, including high-quality responses to user queries from other
large language models, as well as question-answering datasets and human
feedback datasets. Koala has been specifically trained on interaction data
scraped from the web, with a focus on data that includes interaction with
highly capable closed-source models such as ChatGPT. The resulting model,
Koala-13B, demonstrates competitive performance to existing models based on
human evaluation of real-world user prompts.
The
results suggest that using high-quality datasets can overcome some of the
limitations of smaller models and may even match the capabilities of large
closed-source models in the future. The research team recommends that the
community should prioritize curating high-quality datasets, as this may enable
the creation of safer, more factual, and more capable models than simply
increasing the size of existing systems.
One
of the interesting aspects of Koala was the data sources used for training. The
fine-tuning datasets include data curated from ChatGPT dialogs. The fine-tuning
strategy included the following datasets:
·
ShareGPT: Around 60K dialogues shared by users on ShareGPT
were collected through public APIs. The team deduplicated to the user-query level to ensure data quality and removed non-English conversations. The
resulting dataset comprises approximately 30K examples.
·
HC3: The
team used the human and ChatGPT responses from the HC3 English dataset, which
includes roughly 60K human answers and 27K ChatGPT answers for approximately
24K questions. This results in a total of around 87K question-answer examples.
·
OIG: A subset of components from the Open Instruction
Generalist dataset curated by LAION was used, including the
grade-school-math-instructions, the poetry-to-songs, and the
plot-screenplay-books-dialogue datasets. The selected subset results in a total
of around 30K examples.
·
Stanford Alpaca: The team included the dataset used
to train the Stanford Alpaca model, which contains approximately 52K examples
generated by OpenAI’s text-DaVinci-003 through the self-instruct process. It is
worth noting that HC3, OIG, and Alpaca datasets are single-turn question
answering while the ShareGPT dataset is dialogue conversations.
·
Anthropic HH: The team utilized the Anthropic HH
dataset, which includes around 160K human-rated examples. Each example consists
of a pair of responses from a chatbot, one of which is preferred by humans. The
dataset provides both capabilities and additional safety protections for the
model.
·
OpenAI WebGPT: The OpenAI WebGPT dataset includes
approximately 20K comparisons where each example comprises a question, a pair
of model answers, and metadata. The answers are rated by humans with a
preference score.
·
OpenAI Summarization: The OpenAI summarization dataset
contains approximately 93K examples, each consisting of human feedback regarding the summarizations generated by a model. Human evaluators
chose the superior summary from two options.
A
comparison between Koala, ChatGPT, and open-source models like Alpaca can be
seen in the following matrix:
Datasets
and Training
A
primary obstacle in building dialogue models is curating training data.
Prominent chat models, including ChatGPT, Bard, Bing Chat, and Claude use
proprietary datasets built using significant amounts of human annotation. To
construct Koala, we curated our training set by gathering dialogue data from
the web and public datasets. Part of this data includes dialogues with large
language models (e.g., ChatGPT) that users have posted online.
Rather
than maximizing quantity by scraping as much web data as
possible, we focus on collecting a small high-quality dataset. We
use public datasets for question answering, human feedback (responses rated
both positively and negatively), and dialogues with existing language
models.
Limitations
and Safety
Like
other language models, Koala has limitations and can be harmful when misused.
We observe that Koala can hallucinate and generate non-factual responses with a
highly confident tone, likely due to the fine-tuning dialogue.
Perhaps an unfortunate implication of this is that smaller models inherit the
confident style of larger language models before they inherit
the same level of factuality—if true, this is a limitation that is important to
study in future work. When misused, the hallucinated responses from Koala can
potentially facilitate the spread of misinformation, spam, and other content.
Koalas
can hallucinate inaccurate information in a confident and convincing tone.
Beyond hallucinations, Koala shares deficiencies with other chatbot language
models. Some of which include:
- Biases and Stereotypes: Our model
will inherit biases from the dialogue data it was trained on, possibly
perpetuating harmful stereotypes, discrimination, and other harms.
- Lack of Common Sense: While large
language models can generate text that appears to be coherent and
grammatically correct, they often lack common sense knowledge that humans
take for granted. This can lead to nonsensical or inappropriate responses.
- Limited Understanding: Large language
models can struggle to understand the context and nuances of dialogue.
They can also have difficulty identifying sarcasm or irony, which can lead
to misunderstandings.
To
address the safety implications of Koala, we included adversarial prompts in
the dataset from ShareGPT and Anthropic HH to make the model more robust and
harmless. To further mitigate potential misuse, we deploy OpenAI’s content
moderation filter in our online demo to flag and remove unsafe content. We will
be cautious about the safety of Koala, and we are committed to performing
further safety evaluations of it while also monitoring our interactive
demo.
Conclusion
We
hope that the Koala model will serve as a useful platform for future academic
research on large language models: the model is capable enough to exhibit many
of the capabilities that we associate with modern LLMs while being small enough
to be finetuned or utilized with more limited computing.
Reference
1. https://bair.berkeley.edu/blog/2023/04/03/koala/
HITANSH LAKKAD
Business Analytics intern at
Hunnarvi Technologies Pvt Ltd in collaboration with Nanobi Analytics.
VIEWS ARE PERSONAL
#koala #LLM #AI
#businessanalytics #technology #hunnarvi #nanobi #isme
Comments
Post a Comment