Kyle Avery
@kyleavery_
Followers
4K
Following
4K
Media
128
Statuses
990
i just posted another verifier to the Environments Hub, this time for @picoctf
https://t.co/J76KlNaj9z
app.primeintellect.ai
Verifier for PicoCTF challenges
2
1
22
i ended up changing the original benchmark code a bit: - no more vision option, PDF reports are always presented as text - PDF reports were converted to markdown using OCR + manual review - Hybrid Analysis reports are now indented JSON instead of a big blob
1
0
3
this was just a small project to make the original work from CrowdStrike and Meta more accessible https://t.co/8iWU5kbRmy
1
0
3
i decided to make a nicer dataset to use with my Verifiers implementation of both CyberSOCEval benchmarks Dataset: https://t.co/jX2Hc6kGNU Verifier:
app.primeintellect.ai
Verifiers implementation of CyberSOCEval
why does cybersoceval download PDF from wayback or the source and convert them to text/images for every user đź’€
2
3
19
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other
2
0
13
1
0
19
i will be doing this once and uploading to HF (once wayback machine stops rate limiting me)
0
0
1
why does cybersoceval download PDF from wayback or the source and convert them to text/images for every user đź’€
1
0
1
Please make sure you are only drinking as much water as you REALLY need. We need that for the datacenters. If you’re thirsty, grok is thirsty too.
114
6K
67K
if you are building a framework for research/experimentation, i think it’s important to avoid too much abstraction. even if you want to have fancy classes, i need easy visibility into the underlying prompts, tool calls, batching, etc. i’ve been burned too many times to trust your
1
1
15
curious about RL? learn to train an LLM with me in a couple weeks!
Go beyond superficial usage of #LLMs and #AI in this free training from Cobalt Strike and Outflank experts. Gain practical experience to architect AI-powered attack chains and navigate AI assisted adversary simulation. Register now! https://t.co/sEkgkbungA
0
2
27
Been a long time since I've written something for my blog. Recently got inspired to break down how a very basic evasion attack on a machine learning model might work. Check it out https://t.co/JOnvSPztev
steve-s.gitbook.io
An example evasion attack against (probably) the worst machine learning classifier of all time
2
37
122
Listen to me now: Never, never take these computer people at face value. As soon as they build a machine of overwhelming might they are going to send it against you and everyone you love. Not to kill you (unless you live in a certain part of the world where it is legal to kill as
0
0
3
there would have been ten circles if dante knew about short form video app developers
This is the Sora app, powered by Sora 2. Inside the app, you can create, remix, and bring yourself or your friends into the scene through cameos—all within a customizable feed designed just for Sora videos. See inside the Sora app👇
2
1
8
guess i’ll try lora again
LoRA makes fine-tuning more accessible, but it's unclear how it compares to full fine-tuning. We find that the performance often matches closely---more often than you might expect. In our latest Connectionism post, we share our experimental results and recommendations for LoRA.
0
0
5