Llama-SEA-LION-v3-70B

Introduction

Our Llama-SEA-LION-v3-70B models have been continued pre-trained on top of Llama 3.1 70B Instruct that is 70 billion parameters in size. Similar to our Llama-SEA-LION-v3-8B model, our Llama-SEA-LION-v3-70B also has a context length of 128K tokens, making them our SEA-LION models with the longest context length to date.

Llama-SEA-LION-v3-70B was trained on data comprised of approximately 200B tokens across 11 SEA languages: Burmese, Chinese, English, Filipino, Indonesia, Khmer, Lao, Malay, Tamil, Thai and Vietnamese.

Llama-SEA-LION-v3-70B-IT was fine-tuned in two stages on approximately 12.3M English instruction-completion pairs alongside a pool of 4.5M Southeast Asian instruction-completion pairs from SEA languages such as Indonesian, Javanese, Sundanese, Tamil, Thai and Vietnamese.

At a glance:

• Model type: Decoder

• Tokenizer: Default tokenizer used in Llama 3.1 70B Instruct

• Available Formats:

  • Base (Llama-SEA-LION-v3-70B)

  • Instruct (Llama-SEA-LION-v3-70B-IT)

  • GGUF (Llama-SEA-LION-v3-70B-IT-GGUF)

• Languages supported:

  1. Burmese

  2. Chinese

  3. English

  4. Filipino

  5. Indonesia

  6. Javanese (Instruct/GGUF only)

  7. Khmer

  8. Lao

  9. Malay

  10. Sundanese (Instruct/GGUF only)

  11. Tamil

  12. Thai

  13. Vietnamese

• License: Llama3.1 Community License

Llama-SEA-LION-v3-70B

Training Infrastructure

Llama-SEA-LION-v3-70B was trained in two stages using MosaicML Composer on the following hardware:

Stage	Training Details	Llama-SEA-LION-v3-70B
First Stage	AWS p5e.48xlarge	8 instances
	Nvidia H200 140GB GPU	64
	Training Duration	200 hrs (step 0 - 9000)
Second Stage	SingTel HGX-100	16 instances
	Nvidia H100 80GB GPU	128
	Training Duration	495 hrs (step 9000 - 47684)

Configuration

HyperParameter	Llama-SEA-LION-v3-70B
Precision	bfloat16
Optimizer	decoupled_adamw
Scheduler	weight_stable_decay
Learning Rate	1.0e-5
Global Batch Size	512

Tokenizer

For tokenisation, the model employs the default tokenizer used in Llama 3.1 70B Instruct.

Training Data

Llama-SEA-LION-v3-70B base model was continued pre-trained on 200B tokens of the following data:

Language	Source	Total Tokens (B)	Percentage (%)	Total percentage (%)
Code	Stackv2	40	20	20
English	Dolma	37.5	18.75	25
	Fineweb-Edu	7.5	3.75
	Others	5	2.5
Chinese	SEA-LION Pile v1	12	6	13
	Others	14	7
Vietnamese	SEA-LION Pile v1	8.4	4.2	13
	VinBigData	16	8
	Others	1.6	0.8
Indonesian	SEA-LION Pile v1	7	3.5	13
	SEA-LION Pile v2	7	3.5
	Others	12	6
Thai	SEA-LION Pile v1	10.7	5.35	10
	WangChanBERTa	8.5	4.25
	Others	0.8	0.4
Filipino - Malay - Tamil	SEA-LION Pile v1, AI4Bharat Sangraha	4.28	2.14	3
	Others	1.72	0.86
Khmer - Lao - Burmese	SEA-LION Pile v1	5.2	2.6	3
	Others	0.8	0.4

Note:

• All token counts are counted using Llama 3.1 70B Instruct tokenizer

• SEA-LION Pile v1 is processed from Common Crawl WET, which is published here. The cutoff date of this version is September 2020.
• SEA-LION Pile v2 is processed from Common Crawl WARC from October 2020 to April 2024.

• Tamil data from Sangraha is published here. The paper can be found here.
• Tamil news is sourced with permission from Seithi

Benchmark Performance

We evaluated Llama-SEA-LION-v3-70B base model on general language capabilities and constraint-following behaviour.

General Language Capabilities and Constraint-following Behaviour

For the evaluation of general language capabilities, we employed the SEA-HELM (also known as BHASA) evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal) and Natural Language Inference (NLI).

Note: SEA-HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.

The evaluation was done five-shot with native prompts on a sample of 100-1000 instances for each dataset.

Following the implementation of IFEval in OpenLLM leaderboard, we also implement SEA-IFEval to provide a comparison of the ability of the model to follow specific constraints in English and in SEA languages.

SEA-IFEval

Based on IFEval, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.

SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).

For more details on Llama-SEA-LION-v3-70B base benchmark performance, please refer to the SEA-HELM leaderboard, https://leaderboard.sea-lion.ai/.

Llama-SEA-LION-v3-70B-IT

Fine-Tuning Methodology

Llama-SEA-LION-v3-70B-IT is a multilingual instruction-following model that has been tuned using a combination of a full parameter fine-tune, on-policy alignment, and model merges of the best performing checkpoints. The training process for fine-tuning was approximately 3200 GPU hours, on a single node of 8x H100-80GB GPUs.

Fine-Tuning Data

Llama-SEA-LION-v3-70B-IT was trained on a wide range of synthetic instructions, alongside publicly available instructions hand-curated by the team with the assistance of native speakers. In addition, special care was taken to ensure that the datasets used had commercially permissive licenses through verification with the original data sources.

Benchmark Performance

We evaluated Llama-SEA-LION-v3-70B-IT on both general language capabilities and instruction-following capabilities.

General Language Capabilities

For the evaluation of general language capabilities, we employed the SEA-HELM (also known as BHASA) evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal) and Natural Language Inference (NLI).

Note: SEA-HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.

The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.

Instruction-following Capabilities

Since Llama-SEA-LION-v3-70B-IT is an instruction-following model, we also evaluated it on instruction-following capabilities with two datasets, SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench).

As these two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.

SEA-IFEval

SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).

SEA-MTBench

SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs. We use gpt-4-1106-preview as the judge model and compare against gpt-3.5-turbo-0125 as the baseline model. The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction). A tie is given a score of 0.5.

For more details on Llama-SEA-LION-v3-70B-IT benchmark performance, please refer to the SEA-HELM leaderboard, https://leaderboard.sea-lion.ai/.

Llama-SEA-LION-v3-70B-IT-GGUF

The following quantized GGUF formats of our Llama-SEA-LION-v3-70B-IT model are available:

• Llama-SEA-LION-v3-70B-IT-F16

• Llama-SEA-LION-v3-70B-IT-Q2_K

• Llama-SEA-LION-v3-70B-IT-Q3_K_M

• Llama-SEA-LION-v3-70B-IT-Q4_0

• Llama-SEA-LION-v3-70B-IT-Q4_K_M

• Llama-SEA-LION-v3-70B-IT-Q5_0

• Llama-SEA-LION-v3-70B-IT-Q5_K_M

• Llama-SEA-LION-v3-70B-IT-Q6_K

• Llama-SEA-LION-v3-70B-IT-Q8_0

Please refer to our Download the Model(s) section for more details on how to access them.

Download the Model(s)

Llama-SEA-LION-v3-70B models are available for download via the following channels:

HuggingFace SEA-LION v3 Collection

Model	Download
Llama-SEA-LION-v3-70B	HuggingFace
Llama-SEA-LION-v3-70B-IT	HuggingFace
Llama-SEA-LION-v3-70B-IT-GGUF	HuggingFace, Ollama

Usage

Llama-SEA-LION-v3-70B-IT can be run using the 🤗 Transformers library

import transformers
import torch

model_id = "aisingapore/Llama-SEA-LION-v3-70B-IT"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)
messages = [
    {"role": "user", "content": "Apa sentimen dari kalimat berikut ini?\nKalimat: Buku ini sangat membosankan.\nJawaban: "},
]

outputs = pipeline(
    messages,
    max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Disclaimer

It is important for users to be aware that our models exhibits certain limitations that warrant consideration:

1. The model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies in its reasoning.

2. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.

References

Thai Pre-Training Data Reference

@misc{lowphansirikul2021wangchanberta,
    title={WangchanBERTa: Pretraining transformer-based Thai Language Models},
    author={Lalita Lowphansirikul and Charin Polpanumas and Nawat Jantrakulchai and Sarana Nutanong},
    year={2021},
    eprint={2101.09635},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}