Hugging Face's logo
Dipankar Sarkar's picture
🤝 Open to Collab

Dipankar Sarkar PRO

dipankarsarkar
7 1657 329
b4ph's profile picture WANG-CC's profile picture chaoliangUNSW's profile picture
·
https://www.dipankar.cc

AI & ML interests

Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.

Recent Activity

reacted to comgen42's post with 🔥 1 day ago
Kodiak v0.4 is out: cortex-agent-llc/kodiak-v0.4-1b (plus an accuracy mode that averages three runs). This release came out of public feedback. After v0.3, someone showed that two of its skills were answering from keywords instead of reading the case. So we built tests a keyword shortcut can't pass: the same case twice, with one detail changed so the right answer flips. Refund eligibility is now fixed: it gets both versions right 79% of the time, up from 22%. The push-to-main-with-failing-tests case that started this now gets "ask the user first". Agent step safety improved but isn't fixed, and the model card says plainly not to use it as a safety control. Sarcasm still shows no signal on real tweets, and accuracy on brand-new kinds of task hasn't moved. That's the big problem for the next version. Now a pause. I'm shutting the training box down for two weeks while I travel in Turkey. I'll be walking through ancient ruins: places where people built things that lasted thousands of years without any of our tools. I'm hoping that does what travel usually does for me, shakes loose some ideas. I'll be thinking about how to teach Kodiak to handle tasks it has never seen, how to run my publishing company better, and where agentic automation actually earns its keep. No training while I'm gone. When I'm back, Kodiak gets the ideas. Every number, including the misses, is in the public build log: github.com/grizzlypeaksoftware/kodiak
reacted to appvoid's post with 🔥 1 day ago
Byte-level is surprisingly powerful, don't believe me? have a look at this model https://huggingface.co/dotlabs/void.1
reacted to Twu31's post with 🔥 1 day ago
Can a Jev-style decision interface work on EEG when the questions are asked in words? In our tests, only for questions the model was trained on. A small head over frozen EEG features (encode a window once, answer several typed questions about it) was asked each question by a question number, a label template or a description. On SSVEP (BETA, 70 people), asking a seen question with a label template instead of its number cost −5.96 pp of balanced accuracy on a plain spectrum (interval −6.84 to −5.08 pp), the 2 pp margin not met. On flicker frequencies the head was never trained on, it reached 28.7%, where training-free CCA reached 80.9% (chance 12.5%). Results and limits: https://bci.report/topics/questions-in-language/ What "Jev-style" means here: https://bci.report/jev-style/ Query the numbers over MCP: https://huggingface.co/spaces/Twu31/bci-report-explorer Related write-up: https://huggingface.co/blog/Twu31/one-eeg-encoder-several-questions
View all activity

Organizations

Skelf Research's profile picture Neul Labs's profile picture Incredlabs's profile picture Cognisoc's profile picture