Log inSign up
Log inSign up
MekkCyber
White Circle
232 posts
MekkCyber profile banner
@mekkcyber

MekkCyber

White Circle
@mekkcyber
∇ llm research @⚪ | prev @🤗 | msc @parissaclay
Paris
Joined May 2024
908 Following
860 Followers
RepliesRepliesRepostsRepostsMediaMedia

Log in or sign up for X

See what’s happening and join the conversation

Continue with phone
or
Log in with username or email
Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
  • @mekkcyber
    MekkCyber
    White Circle
    @mekkcyber
    Sep 21
    The wait is finally over 🚀 !! Today we're finally shipping Halo A new training framework built to make training simpler, faster, and more reliable. Go try it at: github.com/whitecircle/ha… ∇
    @whitecircle
    White Circle
    @whitecircle
    Sep 21
    Introducing Halo, the best framework for post-training of open-source models. Halo delivers up to 2.8x the throughput of stock TRL with less peak memory, while models stay in their native HuggingFace format. Star us on GitHub: github.com/whitecircle/ha…
    00:00
    5
  • @mekkcyber
    MekkCyber
    White Circle
    @mekkcyber
    Sep 18
    Interesting that jev believes Alibaba trained it ∇
    1
  • @mekkcyber
    MekkCyber
    White Circle
    @mekkcyber
    Sep 17
    Btw trl wasn’t enough for the runs we wanted to do. we kept the same hf models and made them train faster with less memory (and yes, without megatron!). Added bonus: (Less LOCs)/(model integration) More soon! ∇
    1
  • @mekkcyber
    MekkCyber
    White Circle
    @mekkcyber
    Sep 16
    can't tell you what the white or grey bars are yet. can tell you the pink one is TRL. been cooking this for months. more soon. ∇
    3
  • @mekkcyber
    MekkCyber
    White Circle
    @mekkcyber
    Aug 18
    Very cool blog from Anthropic about failure modes in multi-agent systems: TL;DR: 1 - Early models like Sonnet 4.6 and Opus 4.6 coordinate poorly, with lots of conflicts. Newer models like Mythos and Opus 4.8 mostly “solve” this by avoiding coordination and sticking to
    5