SeaLLM-7B-v2 - Large Language Models for Southeast Asia
BIG NEWS: SeaLLM3 is released with state-of-the-art performance of diverse tasks while specifically enhanced to be more trustworthy. Please consider using that latest model version.
Technical Blog 🤗 Tech Memo 🤗 DEMO Github Technical Report
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
- LM-studio:
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
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
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
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.
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},
}
- Downloads last month
- 2