--- license: llama2 inference: false ---
🤗 Tech Memo 🤗 DEMO Github
We introduce SeaLLM - a family of language models optimized for Southeast Asian (SEA) languages. The SeaLLM-base models (to be released) were pre-trained from [Llama-2](https://huggingface.co/meta-llama/Llama-2-13b-hf), on a tailored publicly-available dataset, which comprises mainly Vietnamese 🇻🇳, Indonesian 🇮🇩 and Thai 🇹🇠texts, along with those in English 🇬🇧 and Chinese 🇨🇳. The pre-training stage involves multiple stages with dynamic data control to preserve the original knowledge base of Llama-2 while gaining new abilities in SEA languages. The [SeaLLM-chat](https://huggingface.co/spaces/SeaLLMs/SeaLLM-Chat-13b) model underwent supervised finetuning (SFT) on a mix of public instruction data (e.g. [OpenORCA](https://huggingface.co/datasets/Open-Orca/OpenOrca)) and a small internally-collected amount of natural queries from SEA native speakers, which **adapt to the local cultural norms, customs, styles and laws in these regions**, as well as other SFT enhancement techniques (to be revealed later). Our customized SFT process helps enhance our models' ability to understand, respond and serve communities whose languages are often neglected by previous [English-dominant LLMs](https://arxiv.org/abs/2307.09288), while outperforming existing polyglot LLMs, like [BLOOM](https://arxiv.org/abs/2211.05100) or [PolyLM](https://arxiv.org/pdf/2307.06018.pdf). Our [first released SeaLLM](https://huggingface.co/spaces/SeaLLMs/SeaLLM-Chat-13b) supports Vietnamese 🇻🇳, Indonesian 🇮🇩 and Thai 🇹ðŸ‡. Future verions endeavor to cover all languages spoken in Southeast Asia. - DEMO: [SeaLLMs/SeaLLM-Chat-13b](https://huggingface.co/spaces/SeaLLMs/SeaLLM-Chat-13b) - Model weights: To be released. - Technical report: To be released.> **Disclaimer**: > We must note that even though the weights, codes and demos are released in an open manner, similar to other pre-trained language models, and despite our best effort in red teaming and safety finetuning and enforcement, our models come with potential risks influenced by complex factors, including but not limited to over-diversified, inaccurate, misleading or potentially harmful generation. > Developers and stakeholders should perform their own red teaming and provide related security measures before deployment, and they must abide by and comply with local governance and regulations. > In no event shall the authors be held liable for any claim, damages, or other liability arising from the use of the released weights, codes or demos. > The logo was generated by DALL-E 3. The following sections summarize the [Pre-training](#pre-training), [Supervised-Finetuning (SFT)](#supervised-finetuning-sft) and [performance evaluations](#evaluation). ## Pre-training ### Vocabulary Expansion Like many English/Latin-dominant LLMs, Llama-2's BPE tokenizer breaks non-european and non-latin linguistic texts into unsustainably long byte-level sequences that cover much shorter semantic meanings, leading to [degraded performance](https://arxiv.org/abs/2306.11372). For instance, it takes 4.3x more tokens to encode the same sentence in Thai compared to that in English. This leads to the models failing to perform summarization and comprehension tasks without exceeding the context length. Our goal for vocabulary expansion is threefold: (1) the number of newly-added tokens must be minimal and only cover the new languages, (2) the tokens should bring the compression ratios of new languages close to that of English, and (3) minimize the disruption of existing European tokens to preserve Llama-2 knowledge. In the end, we obtain **~11K** new tokens for Vi, Id, Th and Zh to augment the original 32000-token vocabulary. Details of our expansion technique will be revealed in our upcoming technical report. As seen in the below table, our new vocabulary reduce the compression ratio from 4.29 to 1.57 for Thai, meaning it can now encode 2.7x longer Thai text given the same context length. Meanwhile, English is only compressed by 0.3%, thus preserving its integrity. |Language | Llama's ratio | Our ratio | # New tokens | --- | --- | --- | --- | | Vi | 2.91 | 1.2488 | 2304 | Zh | 1.99 | 1.1806 | 3456 | Th | 4.29 | 1.5739 | 1536 | Id | 1.76 | 1.1408 | 3840 | En | 1.00 | 0.9976 ### Pre-training Data **Pending Lixin's** ### Pre-training Strategies We conduct pre-training in 4 different stages. Each stage serves different specific objectives and involves dynamic control of data mixture, both unsupervised and supervised, and data specification and categorization. We also employ a novel sequence construction and masking techniques during these stages. More details are to be provided in the technical report. As our goal is for Llama-2 to learn new languages with the least number tokens and computing resources, we control appropriate data mix of new (Vi, Id & Th) and old (En, Zh) languages so that the new vocabulary and knowledge is trained quickly, while relatively maintaining the performance of the original Llama-2 model and establishing a knowledge bridge between new and existing languages. We pre-train our SeaLLM-base in ~4 weeks on 32gpus, clocking ~150B tokens. ## Supervised Finetuning (SFT) ### SFT Data Our supervised finetuning (SFT) data consists of many categories. The largests of them are public and open-source, such as [OpenORCA](https://huggingface.co/datasets/Open-Orca/OpenOrca) and [Platypus](https://huggingface.co/datasets/garage-bAInd/Open-Platypus). As the aforementioned is monolingual, we employ several established or novel automatic techniques to gather more instruction data for SEA languages. More importantly, we engaged native speakers to collect a small amount of natural queries and responses data, which adapts to the local cultural customs, norms and laws. We also collect country-relevant safety data that covers many culturally and legally sensitive topics in each of these countries, which are often ignored, or even in conflict, with western safety data. Therefore, we believe our models are more local-friendly and abide by local rules to a higher degree. ### SFT Strategies We conduct SFT with a relatively balanced mix of SFT data from different categories. We make use of the system prompt during training, as we found it helps induce a prior which conditions the model to a behavioral distribution that focus on safety and usefulness. ## Evaluation ### Peer Comparison One of the most reliable ways to compare chatbot models is peer comparison. With the help of native speakers, we built an instruction test set that focus on various aspects expected in a user-facing chatbot, namely (1) NLP tasks (e.g. translation & comprehension), (2) Reasoning, (3) Instruction-following and (4) Natural and Informal questions. The test set also covers all languages that we are concerned with. **Pending peer comparison**Terms of Use: By using our released weights, codes and demos, you agree and comply with the following terms and conditions:
- Follow LLama-2 License and Terms of Use.
- Strictly comply with the local regulations where you operate at and not attempt to generate or illicit our models to generate locally and internationally illegal and inappropriate content.