Jiao hongbao's Trash

欢迎来到本频道少见的加更环节关于为什么会有加更的原因是： 没有任何原因。 我算是比较随意的人，很多时候心血来潮会信誓旦旦地做很多计划，但是实际上不怎么会执行。当突然有一件事情我想去做的时候，只要当下没有十分要紧的事，我就会抽空去做；当一个比较要紧的事是我不太喜欢的东西的时候，我就会在大量兴趣爱好事件中穿插很重要的事情去做。 这次也不例外，属于是临时起意的加更环节。 乐子： 适才相戏耳近日，清华大学的研究生招生主页进行了一次更新，当然也没有时隔多久，我前几天才刚刚写过一个blog提到相关事件，详见此处。 那么我今天要说的是： 已经和上次看不一样了捏。不知道是不是看了网上大家的反应，或者是jyj有通知不允许这么早？原因未知，总之现在已经改了，比上次看推迟了半个月，有点难绷，相对来说，某所就显得沉稳很多： 并非乐子： 恐怖刀战这个其实是比上次写周记更早一些（指半天）的信息，我只是...

...

evaluate the model qualitativelywe need to evaluate the model’s quality, and sometimes we need to make further adjustments to model’s architecture, hyperparameters, or datasets. we should perform inference using the same input with the PEFT model. 12345678910111213141516171819202122%%timefrom transformers import set_seedset_seed(seed)index = 5dialogue = dataset['test'][index]['dialogue']summary = dataset['test'][index]['summary']prompt =...

...

it is time to fine-tuning the LLMLLMs excel in deal with normal context task, but may not always align with specific task or domains, so we can fine-tuning the pre-existing model, to improve LLMs’ performance and accurancy on particular field. so I will learn LLM fine-tuning in this blog, and start my learning record. the key step of LLM fine-tuning select a model. to tune the model on my laptop, with 16GB CPU and 6GB graphic memory, I need to select a model that can be tuned successfully on...

dataset pre-processingwe should format the prompt in a way that the model can recognize. we should refer the huggingface model documentation, to check the format we should transform the original dataset to. for Phi-2, it should be: 1234567where the model generates the text after "." . To encourage the model to write more concise answers, you can also try the following QA format using "Instruct: <prompt>\nOutput:"Instruct: Write a detailed analogy between mathematics...

in hyprland, it is a trouble about screenshots and clipboard. so this is the simple way to make a screenshot keybind. download a screenshot application first. 1sudo pacman -S grim #or others then write a keybind in your ~/.config/hypr/hyprland.conf 1bind = $mainmod, PRINT, exec, grim ~/Pictures/ and you can take shots with win + PrtSc play with arch homepageclick here to come back to play with archhomepage.

what I have done today: files link report click ISSCC-ideas click ISSCC-review click TensorCIM-ideas click MulTCIM-reivew click rep-powerpoint click and pdf version and the last version here. for now, I have done all thinkerlab research ability assessment.

The review of MulTCIM was completed this morning. here and thie afternoon I have to complete the innovation about two papers, and 3000 words report and at least 20 pages powerpoint, good luck. I have completed the innovative improvement to the An 88.36TOPS/W Bit-Level-Weight-Compressed LLM Accelerator with Cluster-Aligned INT-FP-GEMMLLM Accelerator with Cluster-Aligned INT-FP-GEMM and Bi-Dimensional Workflow Reformulationand Bi-Dimensional Workflow Reformulation. 23:54, I finish the...

关于为什么想写这篇blogps:...