@phattar4phan
on
LLM
2h 43m 47s logged
Modify training configurations, pipeline and fix some error in src/model/mlp.py and src/model/rope.py
1h 16m logged
#1 - Initial project and some configurations, also add first-version utilizations extractor, such as extracting ram data, networking, disk i/o and cpu also reads available sensors temperatures
@phattar4phan
on
LLM
4h 19m 29s logged
Change the dataset from entirely scratch to Dolci-Instruct-SFT, then also add serializer (src/utils/serialize.py) to serialize the dataset into .txt file and then to train it with spm.SentencepieceTrain() (src/utils/tspm.py) also add bin.py to turn tokenized into token IDs for faster dataloder, also removed unnecessary file
38m 28s logged
Retrained the model as I have added ‘label_smoothing=0.1’ to criterion to lower the probability of image that is jot white blood cell to be classified and show as a type of wbc with high confident score, also removed static glassmorphism header from the site and also added grid table with some clustered gradients as the background to not make it static
Ship
Changes requested
I make an AI model that could classify white blood cell from scratch (one input at a time as I dont have enough compute unit), the challenging parts ware that I'm not really that good at math so yeah need some AI to help coding the model part to ensure as much accuracy as possible, although it's minimal but I tried to make it have the highest possible accuracy, the last trained one got 'Peak' at 99.19% quite good but it results in getting 100% confident score every time. Before testing, you can test it directly in the web, no downloading or installation required unless you want to make it locally (test at orion.phattar4phan.workers.dev), Proud to show that at least I can make an AI model, next one will be small language model with only ~7.8m parameters, just wait!
1h 44m 15s logged
Finale version, removed the 5 models comparison as I dont have time to create all that right now. Only 1 model 1 interference and it’s donee
@phattar4phan
on
LLM
10h 44m 9s logged
#1 - Been on configurating the project, also after configuration I go on getting the dataset, while you can search for them online, I wanna be really type of ‘from scratch’ so in this case I also create the dataset from scratch as well. And all those 10hrs 44m, I’ve been manually entering random wikipedia articles to the code to scrap the texts, The main obstacle that make this so long is that sometimes I have typos, wrong words, or even not found which I’ve to also manually resolve all of them. Thats all but been going on around ~1600 articles
17m 59s logged
Rewrite .tsx to plain html, css, js to fit the mission scope
Ship
Created my website, I use AI for designing and some coding as am not really experienced wiyh .tsx and designing. This portfolio includes about section, skills section, projects, experiences and contact. the tech stack for this site are Vite, Tailwindcss, React, Lucide-react, and TypeScript. Also I'm proud that I can finally have a stunning portfolio website after stardance encouraging missions (this project was part of the mission Personal Site)
24m 7s logged
Modified README.md, added about, sections, quick start, tech stacks and how I made it
Ship
Created my personal website, I personally used AI for designing and some coding formatting, as am not much experienced with .tsx and not an expert in professional designing. This site is deployed on personal.phattar4phan.workers.dev (hosted on cloudflare), I'm proud that with stardance.hackclub, I can finally create a personal portfolio with their encouraging mission. you all could also be copying this design whenever you want, this is the best I've done. Thank you.
37m 19s logged
Created a full web portfolio, with contacts, skills, techs, projects
3h 50m 13s logged
Created new landing page and web app page, also bunch of new objs
32m 39s logged
Modified some configurations and deploy or vercel (try it at orion-cnn.vercel.app)
3h 5m 30s logged
- Trained CNN from scratch model, and more
training logs:
Training Logs
┏━━━━━━━┳━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━━┓
┃ Epoch ┃ Loss ┃ Accuracy ┃ Valid ┃ Duration(s) ┃
┡━━━━━━━╇━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━━┩
│ 1/20 │ 0.8463 │ 66.54% │ 86.70% │ 48.77 │
│ 2/20 │ 0.5111 │ 81.43% │ 92.74% │ 47.21 │
│ 3/20 │ 0.3835 │ 86.39% │ 91.02% │ 47.00 │
│ 4/20 │ 0.3203 │ 88.78% │ 93.51% │ 46.79 │
│ 5/20 │ 0.2671 │ 90.80% │ 93.51% │ 47.19 │
│ 6/20 │ 0.2448 │ 91.36% │ 93.58% │ 47.44 │
│ 7/20 │ 0.2161 │ 92.34% │ 94.11% │ 47.21 │
│ 8/20 │ 0.1961 │ 93.18% │ 95.82% │ 47.16 │
│ 9/20 │ 0.1859 │ 93.48% │ 96.25% │ 47.15 │
│ 10/20 │ 0.1693 │ 93.97% │ 95.72% │ 47.82 │
│ 11/20 │ 0.1592 │ 94.35% │ 94.46% │ 47.51 │
│ 12/20 │ 0.1503 │ 94.60% │ 96.07% │ 47.39 │
│ 13/20 │ 0.1405 │ 95.24% │ 95.26% │ 47.59 │
│ 14/20 │ 0.1356 │ 95.07% │ 96.28% │ 47.66 │
│ 15/20 │ 0.1232 │ 95.79% │ 96.21% │ 48.00 │
│ 16/20 │ 0.1218 │ 95.91% │ 96.81% │ 47.60 │
│ 17/20 │ 0.1195 │ 95.93% │ 98.11% │ 47.72 │
│ 18/20 │ 0.1162 │ 96.10% │ 97.54% │ 48.10 │
│ 19/20 │ 0.1076 │ 96.49% │ 97.58% │ 47.76 │
│ 20/20 │ 0.1054 │ 96.36% │ 97.05% │ 47.81 │
└───────┴────────┴──────────┴────────┴─────────────┘
Best: E:17 | Validation Accuracy: 98.11
Resulted in the best validation accuracy at 98.11% and at epoch 17th, only need to train the remaining 4 models
