Video wird geladen...

Video konnte nicht geladen werden

Zur Startseite

We’re open-sourcing PorTAL, our framework for shared task representations and cross model LoRA adaptation. It now spans from hybrid attention models to multimodal systems including Gemma 4 E2B, Mistral 7B & Thinking Machines' Inkling. Code: ramp-public/portallib Models: Hugging Face /RampPublic

171,845 Aufrufe • vor 2 Monaten •via X (Twitter)

24 Kommentare

Profilbild von Ramp Labs
Ramp Labsvor 2 Monaten

Code: Models + data: PyPI:

Profilbild von Ramp
Rampvor 2 Monaten

@thinkymachines OPEN IT UP

Profilbild von Ben Geist
Ben Geistvor 2 Monaten

@thinkymachines OSS is short for ossome

Profilbild von Ryan Amiri
Ryan Amirivor 2 Monaten

@thinkymachines @miramurati you seeing what I'm seeing?

Profilbild von shevchenko
shevchenkovor 2 Monaten

@thinkymachines You're telling me a portal trained this Lora?

Profilbild von Alex L
Alex Lvor 2 Monaten

@thinkymachines Ramp Labs ♥️ OSS

Profilbild von Terry Chen
Terry Chenvor 2 Monaten

@thinkymachines Is this still the credit card company? 😟

Profilbild von AI Mastery Guide
AI Mastery Guidevor 2 Monaten

@thinkymachines Cross model adaptation, open sourced for everyone.

Profilbild von Jalen Xing
Jalen Xingvor 2 Monaten

@thinkymachines lfg

Profilbild von sandra
sandravor 2 Monaten

@thinkymachines based based based

Profilbild von Marena
Marenavor 2 Monaten

@thinkymachines Shared representations could simplify model adaptation

Profilbild von Shubhayan Srivastava
Shubhayan Srivastavavor 2 Monaten

@thinkymachines Ahh yes the AI lab for finance

Profilbild von Nick Venturi
Nick Venturivor 2 Monaten

@thinkymachines running this tonight instead of sleeping

Profilbild von ghostkey.base.eth
ghostkey.base.ethvor 2 Monaten

@thinkymachines Open source everything.

Profilbild von René Sultan
René Sultanvor 2 Monaten

@thinkymachines @b_geist did you a try a LoRA for you posts?

Profilbild von Steven Cheng
Steven Chengvor 2 Monaten

@thinkymachines Cross-model LoRA adaptation is a huge unlock. Really curious to see how the shared task representations handle the gap between hybrid attention and multimodal systems like Gemma.

Profilbild von Rain Li
Rain Livor 2 Monaten

@thinkymachines huge

Profilbild von Jurly
Jurlyvor 2 Monaten

@thinkymachines This is huge for the open-source ecosystem 🔥

Profilbild von Sebastian Buzdugan
Sebastian Buzduganvor 2 Monaten

@thinkymachines cross-model lora usually breaks on tokenizer drift before attention differences show up

Profilbild von 安叫兽|Bird🕊️ 🔶 BNB
安叫兽|Bird🕊️ 🔶 BNBvor 2 Monaten

@thinkymachines 跨模型 LoRA 这块终于有人系统搞了

Profilbild von Yechan Do
Yechan Dovor 2 Monaten

@thinkymachines Did they pivot to AI company?

Profilbild von buildwithjay
buildwithjayvor 2 Monaten

@thinkymachines PorTAL is interesting because Ramp has production pain behind it. infra launches land better when the use case is not imaginary

Profilbild von Elon Orbit
Elon Orbitvor 2 Monaten

@thinkymachines If PorTAL unifies adapters across model families, my fine-tuning costs collapse. Big if true. Any vision-language benchmarks yet?

Profilbild von Ah Oh
Ah Ohvor 2 Monaten

@thinkymachines Freezing the task representation and core decoder and recalibrating only the aligner is a clean split. What I'd want to know is how far the aligner stretches before it breaks - Mistral 7B to a multimodal base like Gemma 4 E2B is a big jump in representation geometry.

Ähnliche Videos