We introduce SeaLLM-7B-v2, the state-of-the-art multilingual LLM for Southeast Asian (SEA) languages 🇬🇧 🇨🇳 🇻🇳 🇮🇩 🇹🇭 🇲🇾 🇰🇭 🇱🇦 🇲🇲 🇵🇭. It is the most significant upgrade since SeaLLM-13B, with half the size, outperforming performance across diverse multilingual tasks, from world knowledge, math reasoning, instruction following, etc.

Highlights

SeaLLM-7B-v2 achieves the 7B-SOTA on the Zero-shot CoT GSM8K task with 78.2 score and outperforms GPT-3.5 in many GSM8K-translated tasks in SEA languages (🇨🇳 🇻🇳 🇮🇩 🇹🇭) as well as MGSM (🇨🇳 🇹🇭). It also surpasses GPT-3.5 in MATH CoT for Thai 🇹🇭.
It scores competitively against GPT-3.5 in many zero-shot CoT commonsense benchmark, with 82.5, 68.3, 80.9 scores on Arc-C, Winogrande, and Hellaswag.
It achieves 7.54 score on the 🇬🇧 MT-bench, it ranks 3rd place on the leaderboard for 7B category and is the most outperforming multilingual model.
It scores 45.74 on the VMLU benchmark for Vietnamese 🇻🇳, and is the only open-source multilingual model that can be competitive to monolingual models (Vistral-7B) of similar sizes.

Release and DEMO

DEMO: SeaLLMs/SeaLLM-7B.
Technical report: Arxiv: SeaLLMs - Large Language Models for Southeast Asia.
Model weights:
- SeaLLM-7B-v2.
- SeaLLM-7B-v2-gguf.
- SeaLLM-7B-v2-GGUF (thanks Lonestriker). NOTE: use seallm.preset.json to work properly.
Run locally:
- LM-studio:
  - SeaLLM-7B-v2-q4_0 and SeaLLM-7B-v2-q8_0.
  - LM-studio requires this seallm.preset.json to set chat template properly.
- ollama ollama run nxphi47/seallm-7b-v2:q4_0
- MLX for Apple Silicon: mlx-community/SeaLLM-7B-v2-4bit-mlx

Terms of Use and License: By using our released weights, codes, and demos, you agree to and comply with the terms and conditions specified in our SeaLLMs Terms Of Use.

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 efforts in red teaming and safety fine-tuning and enforcement, our models come with potential risks, including but not limited to 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.

What's new since SeaLLM-13B-v1 and SeaLLM-7B-v1?

SeaLLM-7B-v2 is continue-pretrained from Mistral-7B and underwent carefully designed tuning with focus in reasoning.

Evaluation

Zero-shot CoT Multilingual Math Reasoning

SeaLLM-7B-v2 achieves with 78.2 score on the GSM8K with zero-shot CoT reasoning, making it the state of the art in the realm of 7B models. It also outperforms GPT-3.5 in the same GSM8K benchmark as translated into SEA languages (🇨🇳 🇻🇳 🇮🇩 🇹🇭). SeaLLM-7B-v2 also surpasses GPT-3.5 on the Thai-translated MATH benchmark, with 22.4 vs 18.1 scores.

See details on English and translated GSM8K and MATH with zero-shot reasoning

Model	GSM8K en	MATH en	GSM8K zh	MATH zh	GSM8K vi	MATH vi	GSM8K id	MATH id	GSM8K th	MATH th
GPT-3.5	80.8	34.1	48.2	21.5	55	26.5	64.3	26.4	35.8	18.1
Qwen-14B-chat	61.4	18.4	41.6	11.8	33.6	3.6	44.7	8.6	22	6
Vistral-7b-chat	48.2	12.5			48.7	3.1
Qwen1.5-7B-chat	56.8	15.3	40	2.7	37.7	9	36.9	7.7	21.9
SeaLLM-7B-v2	78.2	27.5	53.7	17.6	69.9	23.8	71.5	24.4	59.6	22.4

Baselines were evaluated using their respective chat-template and system prompts (Qwen1.5-7B-chat, Vistral).

Zero-shot MGSM

SeaLLM-7B-v2 also outperforms GPT-3.5 and Qwen-14B on the multilingual MGSM for Zh and Th.

Model	MGSM-Zh	MGSM-Th
ChatGPT (reported)	61.2	47.2
Qwen-14B-chat	59.6	28
SeaLLM-7B-v2	64.8	62.4

Zero-shot Commonsense Reasoning

We compare SeaLLM-7B-v2 with ChatGPT and Mistral-7B-instruct on various zero-shot commonsense benchmarks (Arc-Challenge, Winogrande and Hellaswag). We use the 2-stage technique in (Kojima et al., 2023) to grab the answer. Note that we DID NOT use "Let's think step-by-step" to invoke explicit CoT.

0-shot reasoning	Arc-Challenge	Winogrande	Hellaswag
ChatGPT (reported)	84.6*	66.8*	72.0*
ChatGPT (reproduced)	84.1	63.1	79.5
Mistral-7B-Instruct	68.1	56.4	45.6
Qwen1.5-7B-chat	79.3	59.4	69.3
SeaLLM-7B-v2	82.5	68.3	80.9

Baselines were evaluated using their respective chat-template and system prompts (Qwen1.5-7B-chat, Mistral).

Multilingual World Knowledge

We evaluate models on 3 benchmarks following the recommended default setups: 5-shot MMLU for En, 3-shot M3Exam (M3e) for En, Zh, Vi, Id, Th, and zero-shot VMLU for Vi.

Model	Langs	En MMLU	En M3e	Zh M3e	Vi M3e	Vi VMLU	Id M3e	Th M3e
GPT-3.5	Multi	68.90	75.46	60.20	58.64	46.32	49.27	37.41
Vistral-7B-chat	Mono	56.86	67.00	44.56	54.33	50.03	36.49	25.27
Qwen1.5-7B-chat	Multi	61.00	52.07	81.96	43.38	45.02	24.29	20.25
SeaLLM-7B-v2	Multi	61.89	70.91	55.43	51.15	45.74	42.25	35.52

VMLU reproduce script here. Lm-eval was used to evaluate MMLU. 0-shot VMLU scores for baselines were evaluated using their respective chat-template and system prompts (Qwen1.5-7B-chat).

MT-Bench

On the English MT-bench metric, SeaLLM-7B-v2 achieves 7.54 score on the MT-bench (3rd place on the leaderboard for 7B category), outperforms many 70B models and is arguably the only one that handles 10 SEA languages.

Refer to mt_bench/seallm_7b_v2.jsonl for the MT-bench predictions of SeaLLM-7B-v2, and here to reproduce it.

Model	Access	Langs	MT-Bench
GPT-4-turbo	closed	multi	9.32
GPT-4-0613	closed	multi	9.18
Mixtral-8x7b (46B)	open	multi	8.3
Starling-LM-7B-alpha	open	mono (en)	8.0
OpenChat-3.5-7B	open	mono (en)	7.81
SeaLLM-7B-v2	open	multi (10+)	7.54
Qwen-14B	open	multi	6.96
Llama-2-70B	open	mono (en)	6.86
Mistral-7B-instuct	open	mono (en)	6.84

Sea-Bench

Similar to MT-Bench, Sea-bench is a set of categorized instruction test sets to measure models' ability as an assistant that is specifically focused on 9 SEA languages, including non-Latin low-resource languages.

As shown, the huge improvements come from math-reasoning, reaching GPT-3.5 level of performance.

Refer to sea_bench/seallm_7b_v2.jsonl for the Sea-bench predictions of SeaLLM-7B-v2.

Usage

Instruction format

python

prompt = """<|im_start|>system
You are a helpful assistant.</s><|im_start|>user
Hello world</s><|im_start|>assistant
Hi there, how can I help?</s>"""

# NOTE: previous commit has \n between </s> and <|im_start|>, that was incorrect!
# <|im_start|> is not a special token.
# Transformers chat_template should be consistent with vLLM format below.

# ! ENSURE 1 and only 1 bos `<s>` at the beginning of sequence
print(tokenizer.convert_ids_to_tokens(tokenizer.encode(prompt)))

'<s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'system', '<0x0A>', 'You', '▁are', '▁a', '▁helpful', '▁assistant', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'user', '<0x0A>', 'Hello', '▁world', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'ass', 'istant', '<0x0A>', 'Hi', '▁there', ',', '▁how', '▁can', '▁I', '▁help', '?', '</s>']
"""

Using transformers's chat_template

python

from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" # the device to load the model onto

# use bfloat16 to ensure the best performance.
model = AutoModelForCausalLM.from_pretrained("SeaLLMs/SeaLLM-7B-v2", torch_dtype=torch.bfloat16, device_map=device)
tokenizer = AutoTokenizer.from_pretrained("SeaLLMs/SeaLLM-7B-v2")

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello world"},
    {"role": "assistant", "content": "Hi there, how can I help you today?"},
    {"role": "user", "content": "Explain general relativity in details."}
]

encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
print(tokenizer.convert_ids_to_tokens(encodeds[0]))
# ['<s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'system', '<0x0A>', 'You', '▁are', '▁a', '▁helpful', '▁assistant', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'user', '<0x0A>', 'Hello', '▁world', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'ass', 'istant', '<0x0A>', 'Hi', '▁there', ',', '▁how', '▁can', '▁I', '▁help', '▁you', '▁today', '?', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'user', '<0x0A>', 'Ex', 'plain', '▁general', '▁rel', 'ativity', '▁in', '▁details', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'ass', 'istant', '<0x0A>']

model_inputs = encodeds.to(device)
model.to(device)

generated_ids = model.generate(model_inputs, max_new_tokens=1000, do_sample=True, pad_token_id=tokenizer.pad_token_id)
decoded = tokenizer.batch_decode(generated_ids)
print(decoded[0])

Using vLLM

python

from vllm import LLM, SamplingParams
TURN_TEMPLATE = "<|im_start|>{role}\n{content}</s>"
TURN_PREFIX = "<|im_start|>{role}\n"

# There is no \n between </s> and <|im_start|>.

def seallm_chat_convo_format(conversations, add_assistant_prefix: bool, system_prompt=None):
    # conversations: list of dict with key `role` and `content` (openai format)
    if conversations[0]['role'] != 'system' and system_prompt is not None:
        conversations = [{"role": "system", "content": system_prompt}] + conversations
    text = ''
    for turn_id, turn in enumerate(conversations):
        prompt = TURN_TEMPLATE.format(role=turn['role'], content=turn['content'])
        text += prompt
    if add_assistant_prefix:
        prompt = TURN_PREFIX.format(role='assistant')
        text += prompt    
    return text

sparams = SamplingParams(temperature=0.1, max_tokens=1024, stop=['</s>', '<|im_start|>'])
llm = LLM("SeaLLMs/SeaLLM-7B-v2", dtype="bfloat16")

message = "Explain general relativity in details."
prompt = seallm_chat_convo_format(message, True)
gen = llm.generate(prompt, sampling_params)

print(gen[0].outputs[0].text)

Fine-tuning SeaLLM-7B-v2

Should follow the chat format and accurately mask out source tokens. Here is an example.

python

conversations = [
    {"role": "system", "content": "You are helful assistant."},
    {"role": "user", "content": "Hello world."},
    {"role": "assistant", "content": "Hi there, how can I help?"},
    {"role": "user", "content": "Tell me a joke."},
    {"role": "assistant", "content": "Why don't scientists trust atoms? Because they make up everything."},
]
def seallm_7b_v2_tokenize_multi_turns(tokenizer, conversations, add_assistant_prefix=False):
    """
    Inputs:
        conversations: list of dict following openai format, eg
            conversations = [
                {"role": "system", "content": "You are helful assistant."},
                {"role": "user", "content": "Hello world."},
                {"role": "assistant", "content": "Hi there, how can I help?"},
                {"role": "user", "content": "Tell me a joke."},
                {"role": "assistant", "content": "Why don't scientists trust atoms? Because they make up everything."},
            ]
        add_assistant_prefix: whether to add assistant_prefix, only for inference decoding
    Outputs:
        tokenize_output_sample, {
            "input_ids": ...
            "token_type_ids": 1 if train and 0 if masked out (not train)
        }
    During training, need to create a labels, with masked-out tokens = -100 to avoid loss computations.
        labels = sample['input_ids'].clone()
        labels[sample['token_type_ids'] == 0] = -100
    """
    TURN_TEMPLATE = "<|im_start|>{role}\n{content}</s>"
    TURN_PREFIX = "<|im_start|>{role}\n"
    sample = None
    assistant_prefix_len = None
    for turn_id, turn in enumerate(conversations):
        prompt = TURN_TEMPLATE.format(role=turn['role'], content=turn['content'])
        turn_sample = tokenizer(
            prompt, padding=False, truncation=False, verbose=False, add_special_tokens=False,
            return_token_type_ids=True, 
        )
        if turn['role'] == 'assistant':
            if assistant_prefix_len is None:
                assistant_prefix_len = len(tokenizer.encode(TURN_PREFIX.format(role=turn['role']), add_special_tokens=False))
            turn_sample['token_type_ids'][assistant_prefix_len:] = [1] * (len(turn_sample['input_ids']) - assistant_prefix_len)
        if sample is None:
            sample = turn_sample
        else:
            for k in turn_sample.keys():
                sample[k].extend(turn_sample[k])
    if add_assistant_prefix:
        assistant_prefix_sample = tokenizer(
            TURN_PREFIX.format(role="assistant"), padding=False, truncation=False, verbose=False, add_special_tokens=False,
            return_token_type_ids=True, 
        )
        for k in sample.keys():
            sample[k].extend(assistant_prefix_sample[k])
    if tokenizer.add_bos_token:
        sample['input_ids'] = [tokenizer.bos_token_id] + sample['input_ids']
        sample['attention_mask'] = [1] + sample['attention_mask']
        sample['token_type_ids'] = [sample['token_type_ids'][0]] + sample['token_type_ids']
    return sample

# ! testing
sample = seallm_7b_v2_tokenize_multi_turns(tokenizer, conversations)
print(tokenizer.convert_ids_to_tokens(sample['input_ids']))
print(sample['token_type_ids'])
# ['<s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'system', '<0x0A>', 'You', '▁are', '▁hel', 'ful', '▁assistant', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'user', '<0x0A>', 'Hello', '▁world', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'ass', 'istant', '<0x0A>', 'Hi', '▁there', ',', '▁how', '▁can', '▁I', '▁help', '?', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'user', '<0x0A>', 'Tell', '▁me', '▁a', '▁joke', '.', '</s>', '▁<', '|', 'im', '_', 'start', '|', '>', 'ass', 'istant', '<0x0A>', 'Why', '▁don', "'", 't', '▁scientists', '▁trust', '▁atoms', '?', '▁Because', '▁they', '▁make', '▁up', '▁everything', '.', '</s>']
# [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]

Acknowledgement to Our Linguists

We would like to express our special thanks to our professional and native linguists, Tantong Champaiboon, Nguyen Ngoc Yen Nhi and Tara Devina Putri, who helped build, evaluate, and fact-check our sampled pretraining and SFT dataset as well as evaluating our models across different aspects, especially safety.

Citation

If you find our project useful, we hope you would kindly star our repo and cite our work as follows: Corresponding Author: l.bing@alibaba-inc.com

Author list and order will change!

* and ^ are equal contributions.

@article{damonlpsg2023seallm,
  author = {Xuan-Phi Nguyen*, Wenxuan Zhang*, Xin Li*, Mahani Aljunied*,
            Zhiqiang Hu, Chenhui Shen^, Yew Ken Chia^, Xingxuan Li, Jianyu Wang,
            Qingyu Tan, Liying Cheng, Guanzheng Chen, Yue Deng, Sen Yang,
            Chaoqun Liu, Hang Zhang, Lidong Bing},
  title = {SeaLLMs - Large Language Models for Southeast Asia},
  year = 2023,
  Eprint = {arXiv:2312.00738},
}