For several months, I have benefited greatly from the Unsloth project, primarily because a significant part of my job involves fine-tuning large language models (LLMs). Fine-tuning LLMs is extremely time-consuming; aside from data collection, the biggest time sink is the endless GPU-powered fine-tuning process.
Read More »Note Of Unsloth Accelerate Fine-tuning Open Source Project
The accelerated framework is proposed by Huawei Noah’s Ark Lab, it replaces the small model used in the original speculative decoding with the shallow sub-network of the large model. Additionally, it employs an extra-trained adapter and the model’s own decoding head to generate speculative tokens, which are then verified by the large model. The subsequent operations are quite similar to the original speculative decoding process.
Read More »[Paper Reading] Kangaroo: Lossless Self-Speculative Decoding via Double Early Exiting
Prompt injection attacks are a burgeoning security concern, primarily targeting large language models (LLMs) or other AI-related domains.
Read More »Defense Note Against Prompt Injection Attack
SFTTrainer is a LLM fine-tuning tool provided by HuggingFace team, that can easily adjust many hyper-parameters and config at the fine-tuning task. In the process, response_template is the special string template we need to pass into the tool, any response right by it will be computed the loss.
RAG-based LLM is a well-known architecture in current usage of Large Language Models (LLM). It involves “retrieval” to provide the model with prior knowledge that it lacks during training, enabling the model to answer questions in the context of specific information.
Read More »[Paper Reading] Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection
The wave of large models has been unstoppable since the release of ChatGPT in November 2022. Up to now, the scale of open-source Large Language Models (LLMs) continues to increase, such as LLaMA-2-70B and Falcon-180B, to name a few.
Read More »[Paper Reading] QLoRA: Efficient Finetuning of Quantized LLMs
Paper link: https://arxiv.org/abs/2212.13345
The author of this research work is the renowned figure in the field of deep learning, Geoffrey Hinton, who was originally a researcher at Google Brain when he initially wrote this paper (he left in 2023).
Read More »[Paper Reading] The Forward-Forward Algorithm: Some Preliminary Investigation
Yesterday, I developed a model merging program. This time I have no enough gpu memory to merge the models in only one time, so I tried to merge layer by layer. I found the memory of GPU is easily to release but CPU didn’t.
Read More »[PyTorch] Release GPU / CPU Memory After Delete Model
DPO (Direct Preference Optimization) is a fine-tuning method that want to replace RLHF (Reinforcement Learning from Human Feedback).
Read More »Direct Preference Optimization (DPO) Training Note HuggingFace has published an article stating that the current LLM is best trained according to the ChatML format. In normal case, it will be generated according to three different roles of system, user and assistant. The format is as follows:
Read More »[Solved] Mistral Cannot Generate eos_token `<|im_end|>` After SFTTrainer Fine-tuning
