Google has finally introduced a new type of open-weight generative AI, "Gemma2" (1). Although it had been previously announced, it came out sooner than expected. As shown below, the 27B model boasts an impressive 12th place on the leaderboard, closely rivaling larger models. A technical report (2) is also available, so let's take a look at what kind of evolution has occurred.
LMSYS Chatbot Arena Leaderboard
1. Model Architecture
Gemma2 adopts the familiar decoder-only transformer architecture. It's the same as GPT4. The context window, which indicates the amount of information that can be input and output at once, is 8192 tokens. The model structure is largely the same as Gemma1, but according to the technical report, the following points have been updated:
“We alternate between a local sliding window attention (Beltagy et al., 2020) and global attention (Luong et al., 2015) in every other layer. The sliding window size of local attention layers is set to 4096 tokens, while the span of the global attention layers is set to 8192 tokens.”
Global attentional model (3)
Comparison of full self-attention pattern and other attention patterns (4)
2. Pre-training
Gemma2's training data is as follows:
27B model: 13 trillion tokens, primarily English data
9B model: 8 trillion tokens
2.6B model: 2 trillion tokens
"These tokens come from a variety of data sources, including web documents, code, and science articles. Our models are not multimodal and are not trained for state-of-the-art multilingual capabilitiesthe.”
“same tokenizer as Gemma 1 and Gemini: a SentencePiece tokenizer with split digits, preserved whitespace, and byte-level encodings. The resulting vocabulary has 256k entries."
Knowledge distillation was also adopted for the 9B and 2.6B models. In my opinion, this might be the most evolved point of Gemma2. It's a Google-specific strategy to leverage the advantages of their existing large-scale generative AI to improve the performance of smaller models. The technical report explains in detail: "Given a large model used as a teacher, we learn smaller 9B and 2.6B models by distilling from the probability given by the teacher of each token 𝑥 given its context 𝑥𝑐, i.e., 𝑃𝑇(𝑥 | 𝑥𝑐). More precisely, we minimize the negative log-likelihood between the probabilities from the teacher and the student.
where 𝑃𝑆 is the parameterized probability of the student. In practice, we run inference on the teacher once and store the probabilities. Since the vocabulary has 256k entries, we only store a sampled subset of the teacher probabilities."
3. Post-training
This part uses techniques commonly seen in other generative AIs. According to the technical report, it is implemented in the following process:
“For post-training, we fine-tune our pre-trained models into instruction-tuned models. First, we apply supervised fine-tuning (SFT) on a mix of text-only, English-only synthetic and humangenerated prompt-response pairs. We then apply RLHF on top of these models with the reward model trained on labelled English-only preference data and the policy based on the same prompts as the SFT phase. Finally, we average the models obtained after each phase to improve their overall performance.“
It's noteworthy that knowledge distillation is adopted again. "We run behavioral cloning on synthetic and real prompts, and responses predominantly synthetically generated by the teacher, that is a larger model. We also run distillation from the teacher on the student’s distribution." In the future, knowledge distillation from large models to small models may become common practice. It's exciting to see.
What do you think? Gemma2 seems to be a model with high potential even in small sizes, and it's promising. The 2.6B model is also expected to be released soon. By the way, Google, which created Gemma2, and Meta, which created Llama3 that we covered last time, have been rivals in the open-source world for more than 8 years with "Tensorflow vs PyTorch". It seems that a similar battle has begun in generative AI as well. Next time, I'd like to try various things with the Gemma2 model. Stay tuned!
1) Gemma 2 is now available to researchers and developers, Google, 27 June 2024
2) Gemma 2 technical paper, Google DeepMind, 27 June 2024
3) Effective Approaches to Attention-based Neural Machine Translation, Minh-Thang Luong Hieu Pham Christopher D. Manning Computer Science Department, Stanford University, 20 Sep 2015
4) Longformer: The Long-Document Transformer, Iz Beltagy, Matthew E. Peters, Arman Cohan, Allen Institute for Artificial Intelligence, 2 Dec 2020
5) On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes, Rishabh Agarwal12, Nino Vieillard1, Yongchao Zhou13, Piotr Stanczyk1, Sabela Ramos1, Matthieu Geist1, Olivier Bachem1, 1Google DeepMind, 2Mila, 3University of Toronto, 17 Jan 2024
Notice: ToshiStats Co., Ltd. and I do not accept any responsibility or liability for loss or damage occasioned to any person or property through using materials, instructions, methods, algorithms or ideas contained herein, or acting or refraining from acting as a result of such use. ToshiStats Co., Ltd. and I expressly disclaim all implied warranties, including merchantability or fitness for any particular purpose. There will be no duty on ToshiStats Co., Ltd. and me to correct any errors or defects in the codes and the software.