Alan Tseng

agentlans

AI & ML interests

Small data, boring AI

Recent Activity

updated a dataset about 7 hours ago
agentlans/ai-military-ethics-bibliography
updated a model about 11 hours ago
agentlans/EuroLLM-1.7B-Instruct-literary-analysis
published a model about 12 hours ago
agentlans/EuroLLM-1.7B-Instruct-literary-analysis
View all activity

Organizations

None yet

agentlans's activity

reacted to ProCreations's post with ❤️ 1 day ago
view post
Post
2748
Post of the Day

I’m fine-tuning Qwen 2.5-0.5B to be extremely good at math, using high-quality datasets and some smart training strategies.
The logs are looking really promising so far!

Expected release:
Tomorrow morning?
I’ll post as soon as it’s ready — stay tuned.

If you want faster updates or just wanna chat about it, come join my Discord:
https://discord.gg/EXsug2Ux29
(Heads up: we might ask a couple quick questions when you join — just making sure we keep the server safe.)

Also, check out one of the datasets we’re using:
ProCreations/SimpleMath

This project is also helping shape the future of IntellIte.
The insights and techniques we’re developing here — better dataset curation, fine-tuning tricks, and evaluation methods — will directly contribute to making IntellIte even sharper, faster, and more reliable, especially for math and reasoning tasks.

Big progress ahead. Can’t wait to share it with you all!
  • 1 reply
·
replied to ProCreations's post 1 day ago
view reply

I've done something like that before but I got a model overfitted to memorize multiplication tables. 🥲
Starting with the basics is a good idea. Those LLMs are big and hard to debug. Hope it turns out well!

replied to shekkizh's post 2 days ago
view reply

Yeah that's... very odd.

The issue is that of the model not able to reason both color and text in the tokenized space.

Only explanation I can think of is that the model isn't trained to reason with both colour and text information, especially since most UI are designed to be simple and accessible to the end user. That might explain why the UI agent does so poorly on a visual puzzle like Wordle.

I agree with you that this isn't remotely close to AGI. It'll be interesting to see what results you have if you decide to publish.

replied to ZennyKenny's post 2 days ago
view reply

Releasing military tactics data raises concerns. Unlike game theory datasets, which have lower risk and a wider range of uses and scenarios, military data can be misused without important factors like impact on civilians.

Without these elements, AI models might develop oversimplified or harmful strategies. To balance research value with ethical responsibility, you should consider safeguards or adopt a hybrid approach combining game theory with realistic yet abstracted conflict scenarios.