type
status
date
slug
summary
tags
category
password
icon
Author
Abstract
Job
In August, I focused on fine-tuning the
Qwen2-7b model and evaluating its performance on our private benchmark consisting of over 200 questions and answers. I evaluated various large language models (LLMs) like GPT-4, Gemini 1.5-Pro, and Llama 3-405b on this benchmark to compare their capabilities in areas such as reasoning, coding, and commonsense.While the fine-tuned
Qwen2-7b surpassed some older models over 30B parameters, it couldn't match the performance of Qwen2-7b-instruct. Here are some key learnings from the fine-tuning process:- Using LoRA (Low-Rank Adaptation) was a good choice to reduce the number of trainable parameters.
- The dataset quality is crucial. I tried various methods like evol-instruct, persona-hub, and magpie to generate the instruction dataset. Ensuring high dataset quality through techniques like LLLM-as-a-Judge and human annotation is challenging but important.
- DPO (Differentiable Prompt Optimization) along with SFT (Supervised Fine-Tuning) improved performance compared to SFT alone. Despite its benefits, DPO is rarely covered in online fine-tuning tutorials.
I'm considering writing a blog to document my experience with building the evaluation benchmark, fine-tuning models on it, and improving model performance.
I also explored the power of Flux, a new AI model for generating images. I was amazed by its capabilities and the speed at which it generates high-quality images using LoRA fine-tuning on Replicate.
Personal
On the personal front, I delved into DSPy (Differentiable Sparse Polynomial) and tried using MIPRO optimization to improve model performance.
I also fine-tuned a LoRA model for my girlfriend, generating impressive and realistic images.
To better curate and filter information, I started writing a newsletter to collect interesting content, even though it's still in its early stages.
For writing blogs, I've transitioned from Notion and Obsidian to Cursor. The reasons are:
- Cursor allows working with local files, ensuring data safety and control without relying on the internet. While Obsidian also uses local files, Cursor's built-in AI features are more convenient than installing plugins in Obsidian.
- Notion is expensive, and its AI features are not as advanced as Cursor's. Additionally, Notion's recent blocking of Russian users raises concerns about data control.
I still use Notion Next (a free and open-source alternative) for publishing my blogs, but I prefer writing them in Cursor.
That's a recap of my August activities. Looking forward to the next month!
<ins/>
- Author:Chengsheng Deng
- URL:https://chengshengddeng.com/article/recap-for-August
- Copyright:All articles in this blog, except for special statements, adopt BY-NC-SA agreement. Please indicate the source!
Relate Posts