Thinh Truong retweetledi
i hate ML conference reviewers. i take back everything bad i ever said about ACL. every ACL reviewer i ever got was at least literate
English
Thinh Truong
51 posts
Thinh Truong
@ththinh_
Research Fellow @UniMelb
Melbourne Katılım Ocak 2014
356 Takip Edilen46 Takipçiler
@lrzneedresearch 7557 and rejected too. AC even recommended accepting. What's the point of the review process if PC just dismissing everyone...
English
Thinh Truong retweetledi
Guess: Who is this guy?
"Growing up in rural [Location], [Person's Name] didn't have electricity at home. But he lived near a library, where he read compulsively about great inventions and dreamed of adding to the list. He decided around age 14 that humanity would be helped most by a machine smart enough to be an inventor in its own right—an idea that remains only a dream. But it set [Person's Name] on a path toward pioneering an approach to artificial intelligence that could let software understand the world more the way humans do."
Hint: This is someone who inspired me the most in the last few years.
English
also got interrogated at melbourne airport after 30 hrs on the plane. This sucks so much 🙃.
English
@lixin4ever Thank you for the release! One question out of curiosity: What is the motivation behind training LLM for SEA? I know that some languages like malay and indo share some similarities but they are generally disconnected (different script, morphology, culture, etc.)
English
Latest update 🔥🔥🔥:
SeaLLM-7B-v2.5 (huggingface.co/SeaLLMs/SeaLLM…):
- Much more capable than v2 in Thai (+10% gains on Thai exam)
- Multilingually knowledgeable, the best 7B & open-source model on VMLU (53.3% accuracy)
- Still good at math, commonsense reasoning, and instruction-following
SeaExam Leaderboard (huggingface.co/spaces/SeaLLMs…):
- M3Exam + Translated MMLU for assessing LLMs' capabilities in SEA languages
- 60+ representative multilingual LLMs included
- Unified & reasonable evaluation protocols
Check more details of the model and the leaderboard at: damo-nlp-sg.github.io/SeaLLMs/
AK@_akhaliq
SeaLLMs -- Large Language Models for Southeast Asia paper page: huggingface.co/papers/2312.00… Despite the remarkable achievements of large language models (LLMs) in various tasks, there remains a linguistic bias that favors high-resource languages, such as English, often at the expense of low-resource and regional languages. To address this imbalance, we introduce SeaLLMs, an innovative series of language models that specifically focuses on Southeast Asian (SEA) languages. SeaLLMs are built upon the Llama-2 model and further advanced through continued pre-training with an extended vocabulary, specialized instruction and alignment tuning to better capture the intricacies of regional languages. This allows them to respect and reflect local cultural norms, customs, stylistic preferences, and legal considerations. Our comprehensive evaluation demonstrates that SeaLLM-13b models exhibit superior performance across a wide spectrum of linguistic tasks and assistant-style instruction-following capabilities relative to comparable open-source models. Moreover, they outperform ChatGPT-3.5 in non-Latin languages, such as Thai, Khmer, Lao, and Burmese, by large margins while remaining lightweight and cost-effective to operate.
English
Thinh Truong retweetledi
🚨 New paper on Subword Tokenization 🚨
- umLabeller, a new tool, classifies subword tokenization into morph 🤹 or alien 👽
- alien tokenization 🛸 leads to poorer generalizations than morphological tokenization for 3 downstream tasks. arxiv.org/abs/2404.13292 (1/7)
English
0-shot evaluation is important (if not more meaningful) when evaluating general capabilities of LLM.
(((ل()(ل() 'yoav))))👾@yoavgo
"In other words, the modification are so simple that there is a rule to determine the label (e.g., adding 'X allegedly did Y' doe not entail 'X did Y'). " And yet the massively pretrained models do not capture this! I find it interesting and remarkable.
English
11/ Please have a look at the paper if you find this interesting.
Big thanks to my co-author and supervisors: @YuliaOtmakhova, @karinv, Trevor, @eltimster. Finally, I will be at NAACL (my first in-person conference after almost finishing my PhD). See you in Mexico!
English
10/ We see that models could clearly understand negation despite incorrect tokenization. This could be an interesting phenomenon to look at when discovering LLM interpretability. Also, as English is poor in morphology, we are eager to extend this analysis to other languages.
English
Paper accepted to NAACL 2024 main conference.
arxiv: arxiv.org/abs/2404.02421
In this work, we explore the interaction between two bottlenecks of LLMs: negation and tokenization (quoting @karpathy: "tokenization is the root of suffering").
English