We introduce SeaLLM-7B-v2.5, 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-v2.5 outperforms GPT-3.5 and achieves 7B SOTA on most multilingual knowledge benchmarks for SEA languages (MMLU, M3Exam & VMLU).
It achieves 79.0 and 34.9 on GSM8K and MATH, surpassing GPT-3.5 in MATH.

Release and DEMO

DEMO:
- SeaLLMs/SeaLLM-7B-v2.5.
- SeaLLMs/SeaLLM-7B | SeaLMMM-7B - Experimental multimodal SeaLLM.
Technical report: Arxiv: SeaLLMs - Large Language Models for Southeast Asia.
Model weights:
- SeaLLM-7B-v2.5.
- SeaLLM-7B-v2.5-GGUF.
Run locally:
- LM-studio:
  - SeaLLM-7B-v2.5-q4_0-chatml with ChatML template (<eos> token changed to <|im_end|>)
  - SeaLLM-7B-v2.5-q4_0 - must use SeaLLM-7B-v2.5 chat format.
- MLX for Apple Silicon: SeaLLMs/SeaLLM-7B-v2.5-mlx-quantized
Previous models:
- SeaLLM-7B-v2
- SeaLLM-7B-v1

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-7B-v2?

SeaLLM-7B-v2.5 was built on top of Gemma-7b, and underwent large scale SFT and carefully designed alignment.

Evaluation

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
SailorLM	Multi	52.72	59.76	67.74	50.14	---	39.53	37.73
SeaLLM-7B-v2	Multi	61.89	70.91	55.43	51.15	45.74	42.25	35.52
SeaLLM-7B-v2.5	Multi	64.05	76.87	62.54	63.11	53.30	48.64	46.86

Zero-shot CoT Multilingual Math 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	6.0
Vistral-7b-chat	48.2	12.5			48.7	3.1
Qwen1.5-7B-chat	56.8	15.3	40.0	2.7	37.7	9	36.9	7.7	21.9	4.7
SeaLLM-7B-v2	78.2	27.5	53.7	17.6	69.9	23.8	71.5	24.4	59.6	22.4
SeaLLM-7B-v2.5	78.5	34.9	51.3	22.1	72.3	30.2	71.5	30.1	62.0	28.4

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

Zero-shot MGSM

SeaLLM-7B-v2.5 also outperforms GPT-3.5 and Qwen-14B on the multilingual MGSM for Thai.

Model	MGSM-Zh	MGSM-Th
ChatGPT (reported)	61.2	47.2
Qwen-14B-chat	59.6	28
SeaLLM-7B-v2	64.8	62.4
SeaLLM-7B-v2.5	58.0	64.8

Sea-Bench

Usage

IMPORTANT NOTICE for using the model

<bos> must be at start of prompt, ff your code's tokenizer does not prepend <bos> by default, you MUST prepend into the prompt yourself, otherwise, it would not work!
Repitition penalty (e.g: in llama.cpp, ollama, LM-studio) must be set to 1 , otherwise will lead to degeneration!

Instruction format

python

# ! WARNING, if your code's tokenizer does not prepend <bos> by default,
# You MUST prepend <bos> into the prompt yourself, otherwise, it would not work!

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

# <|im_start|> is not a special token.
# Transformers chat_template should be consistent with vLLM format below.

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

"""

Using transformers's chat_template

Install the latest transformers (>4.40)

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-v2.5", torch_dtype=torch.bfloat16, device_map=device)
tokenizer = AutoTokenizer.from_pretrained("SeaLLMs/SeaLLM-7B-v2.5")

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]))

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}<eos>\n"
TURN_PREFIX = "<|im_start|>{role}\n"

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=['<eos>', '<|im_start|>'])
llm = LLM("SeaLLMs/SeaLLM-7B-v2.5", 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-v2.5

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_v25_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}<eos>\n"
    TURN_PREFIX = "<|im_start|>{role}\n"
    TURN_SUFFIX = "<eos>\n"
    TURN_SUFFIX_TAKE = "<eos>"
    sample = None
    assistant_prefix_len = None
    assistant_suffix_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))
            if assistant_suffix_len is None:
                assistant_suffix_len = (
                    len(tokenizer.encode(TURN_SUFFIX.format(role=turn['role']), add_special_tokens=False)) - 
                    len(tokenizer.encode(TURN_SUFFIX_TAKE, add_special_tokens=False))
                )
            turn_sample['token_type_ids'][assistant_prefix_len:-assistant_suffix_len] = [1] * (len(turn_sample['input_ids']) - assistant_prefix_len - assistant_suffix_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_v25_tokenize_multi_turns(tokenizer, conversations)
tokens = tokenizer.convert_ids_to_tokens(sample['input_ids'])
pairs = [(x, y) for x, y in zip(tokens, sample['token_type_ids'])]
print(pairs)

# source and special tokens is masked out (token_type 0), only assistant with <eos> is trained (token_type 1)
# [('<bos>', 0), ('<', 0), ('|', 0), ..., ('assistant', 0), ('\n', 0), ('Hi', 1), ('▁there', 1), (',', 1), ('▁how', 1), ('▁can', 1), ('▁I', 1), ('▁help', 1), ('?', 1), ('<eos>', 1), ('\n', 0), ('<', 0), ... 

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*, Weiwen Xu, Hou Pong Chan,
            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},
}