Sambalingo Serbian Chat

🚀 SambaLingo-Serbian-Chat

SambaLingo-Serbian-Chat is a human-aligned chat model trained in Serbian and English. It addresses the need for multilingual chat capabilities, leveraging direct preference optimization on a base model adapted to Serbian. This enables more natural and accurate interactions in both Serbian and English.

🚀 Quick Start

Loading Model With Hugging Face

Ensure to set use_fast=False when loading the tokenizer.

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("sambanovasystems/SambaLingo-Serbian-Chat", use_fast=False)
model = AutoModelForCausalLM.from_pretrained("sambanovasystems/SambaLingo-Serbian-Chat", device_map="auto", torch_dtype="auto")

Interacting With Model Pipeline

Again, set use_fast=False when loading the tokenizer.

from transformers import pipeline
pipe = pipeline("text-generation", model="sambanovasystems/SambaLingo-Serbian-Chat", device_map="auto", use_fast=False)
messages = [
                {"role": "user", "content": {YOUR_QUESTION}},
]
prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(prompt)[0]
outputs = outputs["generated_text"]

Suggested Inference Parameters

• Temperature: 0.8
• Repetition penalty: 1.0
• Top-p: 0.9

Prompting Guidelines

Use the following chat template to prompt the model:

<|user|>\n{question}</s>\n<|assistant|>\n

Example Prompt and Generations

<|user|>
Od kada datiraju prva naselja na prostoru danasnjeg Beograda, i kada je grad osnovan?</s>
<|assistant|>
Beograd su naseljavali ljudi još u praistoriji, a prva naselja datiraju iz neolitskih vremena, pre oko 7.000 godina. Međutim, grad Beograd nije osnovan upravo u to vreme, već u srednjem veku, kada su Srbi u 9. veku osnovali srednjovekovni grad Beograd, na mestu današnjeg Kalemegdana. Grad je kasnije pretrpeo razna oсвајања, razaranja i obnavljanja, da bi se na kraju razvio u značajnu evropsku metropolu.

✨ Features

• Multilingual Support: Trained in both Serbian and English, enabling seamless communication in these languages.
• Human Alignment: Utilizes direct preference optimization to provide more human-like responses.
• Based on Strong Foundation: Built on the Llama-2-7b model, adapted to Serbian.

📚 Documentation

Model Description

• Developed by: SambaNova Systems
• Model type: Language Model
• Language(s): Serbian, English
• Finetuned from model: Llama-2-7b
• Try this model: SambaLingo-chat-space
• Paper: SambaLingo: Teaching Large Language Models New Languages
• Blog Post: sambalingo-open-source-language-experts

Training Details

The alignment phase follows the recipe for Zephyr-7B, with two stages: supervised fine-tuning (SFT) and Direct Performance Optimization (DPO).

The SFT phase was conducted on the ultrachat_200k dataset, mixed with the Google-translated version. It was trained for one epoch with a global batch size of 512 and a max sequence length of 2048 tokens. A linear decay learning rate of 2e-5 and 10% warmup were used.

The DPO phase was carried out on the ultrafeedback dataset and cai-conversation-harmless dataset, mixed with 10% of the data Google-translated. It was trained with a global batch size of 32 for three epochs. A linear decay learning rate of 5e-7, 10% warmup, and β = 0.1 as the regularization factor for DPO were employed.

Tokenizer Details

The vocabulary of the base llama model was extended from 32,000 tokens to 57,000 tokens by adding up to 25,000 non-overlapping tokens from the new language.

Evaluation

For evaluation results, refer to our paper: SambaLingo: Teaching Large Language Models New Languages

Uses

Direct Use

The use of this model is governed by the Meta’s Llama 2 Community License Agreement. Review and accept the license before downloading the model weights.

Out-of-Scope Use

SambaLingo should not be used for:

• Mission-critical applications
• Applications involving the safety of others
• Making highly important decisions

Bias, Risks, and Limitations

Like all LLMs, SambaLingo has certain limitations:

• Hallucination: The model may generate responses with plausible but factually incorrect or irrelevant information.
• Code Switching: It might unintentionally switch between languages or dialects within a single response, affecting output coherence and understandability.
• Repetition: The model may produce repetitive phrases or sentences, resulting in less engaging and informative responses.
• Coding and Math: Its performance in generating accurate code or solving complex mathematical problems may be limited.
• Toxicity: The model could inadvertently generate responses with inappropriate or harmful content.

Acknowledgments

We express our sincere gratitude to the open-source AI community. This project would not have been possible without open source. SambaNova supports the open-source community and aims to contribute actively.

Special thanks to the following groups:

• Meta for open-sourcing LLama 2 and the FLORES-200 dataset
• Nguyen et al for open-sourcing the CulturaX dataset
• CohereAI for releasing AYA-101 and open-sourcing a multilingual instruction tuning dataset
• EleutherAI for their open-source evaluation framework
• Hugging Face-H4 team for open-sourcing the zephyr training recipe and alignment handbook repo

Cite SambaLingo

@misc{csaki2024sambalingo,
      title={SambaLingo: Teaching Large Language Models New Languages}, 
      author={Zoltan Csaki and Bo Li and Jonathan Li and Qiantong Xu and Pian Pawakapan and Leon Zhang and Yun Du and Hengyu Zhao and Changran Hu and Urmish Thakker},
      year={2024},
      eprint={2404.05829},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

📄 License

The model uses the Llama 2 license.