FLUX #LoRAs work!

It's crazy

This model looks really good and the post-training recipe for SFT models combines a bunch of cool tricks that the community has developed over the past year: - Filter a large SFT corpus for quality x difficulty (similar to Llama3) - Use the Spectrum method from Cognitive Computations

Seguimos con la sesión de tarde con Manu Romero , que nos introducirá en el lenguaje natural, y nos hará un recorrido por el Hub de Hugging Face

Esta tarde en nuestro curso #IAenUNIA haciendo fine tuning y cuantizando un modelo LLM (Qlora) con Manu Romero

Phi goes MoE! Microsoft just released Phi-3.5-MoE a 42B parameter MoE built upon datasets used for Phi-3. Phi-3.5 MoE outperforms bigger models in reasoning capability and is only behind GPT-4o-mini. 👀 TL;DR 🧮 42B parameters with 6.6B activated during generation 👨‍🏫  16

Hoy, en #IAenUNIA, interesantísima sesión sobre cuantización de modelos de lenguaje de gran escala y RAGs, con Manu Romero y María Grandury SomosNLP Universidad Intern. de Andalucía

Small talk is more difficult than expected for chat models, so we built a dataset to fix this! When using standard SFT datasets like Magpie or WebInstruct models often still spectacularly fail when you just greet them.

Continuous self-critique/review in agents (LLMs) is akin to a `while True` loop. It's as if they always have something to say (or improve) and can end with very weird results.

Service LLMs greatly reduce the barrier to entry for many applications that want to endow cool things. However, they impose internet access and privacy concerns. We present LlamaDuo, a simple pipeline that mimics a service LLM on SPECIFIC TASKS through a small LM in crisis

Our solution rocks!

FLUX.1-schnell OpenVINO support github: github.com/rupeshs/fastsd…

Cool release!🚀

Forgot to make it public: huggingface.co/mrm8488/multil… Matryoshka Embeddings model fine-tuned for better performance on Spanish texts

This is a “good enough” way to run *any* company. Founder Mode is N=1

OpenAI’s Strawberry, LM self-talk, inference scaling laws, and spending more on inference Whether or not scaling works, we should spend more on inference. interconnects.ai/p/openai-straw…

In these times when we know data quality is vital to creating better LLMs, I've fine-tuned a set of SoTA Embeddings models on the WebInstruct dataset to help with this kind of task. Hugging Face collection: huggingface.co/collections/mr…