AI Pulse by Inblix

AI Research

Cutting-edge AI research papers, breakthroughs, and academic developments.

601 articles

String-matching metrics are failing VQA models—even when answers are right — Inblix summary
Research

String-matching metrics are failing VQA models—even when answers are right

Hugging Face Blog · Jul 25, 2024 · 2 min read

The way we score visual question answering models is quietly falling apart. On Docmatix, a synthetic document VQA datas...

Llama 3.1 lands: 405B parameters, 128K context, and a license that lets you distill — Inblix summary
Research

Llama 3.1 lands: 405B parameters, 128K context, and a license that lets you distill

Hugging Face Blog · Jul 23, 2024 · 2 min read

Meta just dropped Llama 3.1, and the headline number is absurd: 405 billion parameters. That's the largest dense open-w...

Mistral 7B now runs on Macs in under 4GB memory via Apple's Core ML update — Inblix summary
Research

Mistral 7B now runs on Macs in under 4GB memory via Apple's Core ML update

Hugging Face Blog · Jul 22, 2024 · 2 min read

Apple's WWDC 24 Core ML updates make running a 7-billion-parameter model on consumer hardware surprisingly practical. T...

Docmatix drops 9.5M QA pairs to close the DocVQA gap with closed models — Inblix summary
Research

Docmatix drops 9.5M QA pairs to close the DocVQA gap with closed models

Hugging Face Blog · Jul 18, 2024 · 3 min read

The team behind Idefics2 just released Docmatix, a document visual question answering dataset that makes the previous s...

Hugging Face's TGI now serves 30 LoRA models from a single GPU deployment — Inblix summary
Research

Hugging Face's TGI now serves 30 LoRA models from a single GPU deployment

Hugging Face Blog · Jul 18, 2024 · 2 min read

Hugging Face just solved one of the most annoying problems in LLM deployment: paying for multiple GPUs when you need mu...

Argilla 2.0's Docs Chatbot: Fine-Tuned Embeddings From Synthetic Data — Inblix summary
Research

Argilla 2.0's Docs Chatbot: Fine-Tuned Embeddings From Synthetic Data

Hugging Face Blog · Jul 16, 2024 · 2 min read

The team behind Argilla has open-sourced a practical blueprint for building documentation chatbots that actually unders...

Hugging Face's SmolLM hits above its weight with 1.7B params — Inblix summary
Research

Hugging Face's SmolLM hits above its weight with 1.7B params

Hugging Face Blog · Jul 16, 2024 · 2 min read

Hugging Face just dropped SmolLM, a family of three small language models ranging from 135M to 1.7B parameters, and the...

A 7B Model Just Beat Everyone at Math Olympiad Problems. Here's the Recipe — Inblix summary
Research

A 7B Model Just Beat Everyone at Math Olympiad Problems. Here's the Recipe

Hugging Face Blog · Jul 11, 2024 · 2 min read

The Numina team has just pulled off something remarkable: a fine-tuned 7-billion-parameter model that out-reasoned much...

Hugging Face Deploys Presidio to Scan Datasets for Leaked PII — Inblix summary
Research

Hugging Face Deploys Presidio to Scan Datasets for Leaked PII

Hugging Face Blog · Jul 10, 2024 · 2 min read

Hugging Face is rolling out an experimental feature that automatically scans datasets on the Hub for personally identif...

Hugging Face TRL now trains vision models with DPO on 83K-row datasets — Inblix summary
Research

Hugging Face TRL now trains vision models with DPO on 83K-row datasets

Hugging Face Blog · Jul 10, 2024 · 2 min read

Hugging Face's TRL library just got a capability upgrade that quietly matters more than most model releases: direct pre...

KerasHub now loads 300K+ Hugging Face models in one line of code — Inblix summary
Research

KerasHub now loads 300K+ Hugging Face models in one line of code

Hugging Face Blog · Jul 10, 2024 · 2 min read

The wall between Hugging Face's massive model collection and KerasHub just came down. Previously, KerasHub users could...

France's CDC bets on open-source RAG to renovate 10,000 schools — Inblix summary
Research

France's CDC bets on open-source RAG to renovate 10,000 schools

Hugging Face Blog · Jul 9, 2024 · 2 min read

France's Banque des Territoires is betting that generative AI can accelerate one of the country's most ambitious enviro...

Google TPUs Hit Hugging Face at $1.37/Hour, Targeting Llama and Mistral Deploys — Inblix summary
Research

Google TPUs Hit Hugging Face at $1.37/Hour, Targeting Llama and Mistral Deploys

Hugging Face Blog · Jul 9, 2024 · 2 min read

Hugging Face developers can now rent Google's custom TPU v5e chips directly through Inference Endpoints and Spaces, wit...

Hugging Face adds 4 Dataset Search filters to tame its 350K+ datasets — Inblix summary
Research

Hugging Face adds 4 Dataset Search filters to tame its 350K+ datasets

Hugging Face Blog · Jul 8, 2024 · 2 min read

Hugging Face just gave its Dataset Hub a serious upgrade, rolling out four new search filters that make finding the rig...

Intel's Gaudi 2 fine-tunes ProtST protein model 2.92x faster than Nvidia A100 — Inblix summary
Research

Intel's Gaudi 2 fine-tunes ProtST protein model 2.92x faster than Nvidia A100

Hugging Face Blog · Jul 3, 2024 · 2 min read

Intel and MILA have re-architected ProtST, the multi-modal protein language model that made waves at ICML 2023, and rel...

Hugging Face's Code Agent Tops GAIA Benchmark, Beating GPT-4 Turbo's 7% — Inblix summary
Research

Hugging Face's Code Agent Tops GAIA Benchmark, Beating GPT-4 Turbo's 7%

Hugging Face Blog · Jul 1, 2024 · 2 min read

Hugging Face engineers decided to stress-test their Transformers Agents library against GAIA, widely considered the mos...

Google's Gemma 2 drops: A 27B open model that beats Llama 3 70B — Inblix summary
Research

Google's Gemma 2 drops: A 27B open model that beats Llama 3 70B

Hugging Face Blog · Jun 27, 2024 · 2 min read

Google just released Gemma 2, and the numbers are frankly uncomfortable for anyone who bought into the bigger-is-better...

XLSCOUT's ParaEmbed 2.0 boosts patent search accuracy by 23% — Inblix summary
Research

XLSCOUT's ParaEmbed 2.0 boosts patent search accuracy by 23%

Hugging Face Blog · Jun 25, 2024 · 2 min read

Patent analysis has always been a slog for AI. Generic models choke on claim language, legal jargon, and the kind of te...

Bad data is quietly breaking AI models — and the fix starts before training — Inblix summary
Research

Bad data is quietly breaking AI models — and the fix starts before training

Hugging Face Blog · Jun 24, 2024 · 2 min read

Everyone talks about model architecture, parameter counts, and compute budgets. Fewer people want to discuss the unglam...

Florence-2 Fine-Tuned on DocVQA Jumps from 0 to 57% Similarity — Inblix summary
Research

Florence-2 Fine-Tuned on DocVQA Jumps from 0 to 57% Similarity

Hugging Face Blog · Jun 24, 2024 · 2 min read

Florence-2 was supposed to handle visual question answering out of the box. That's what Microsoft's paper implied. But...

Hugging Face's open data drive pulled 385 volunteers in days — Inblix summary
Research

Hugging Face's open data drive pulled 385 volunteers in days

Hugging Face Blog · Jun 20, 2024 · 2 min read

Hugging Face's Data Is Better Together initiative started with a simple bet: that open-source AI needs better datasets,...

Prezi leans on Hugging Face experts to sharpen its AI presentation pipeline — Inblix summary
Research

Prezi leans on Hugging Face experts to sharpen its AI presentation pipeline

Hugging Face Blog · Jun 19, 2024 · 3 min read

Prezi, the online presentation platform, has been quietly reworking the machine learning that powers its flagship AI pr...

BigCodeBench stress-tests LLMs with 1,140 tasks and 99% branch coverage — Inblix summary
Research

BigCodeBench stress-tests LLMs with 1,140 tasks and 99% branch coverage

Hugging Face Blog · Jun 18, 2024 · 2 min read

The AI community has been stuck with benchmarks that either skew too academic or lean too niche. HumanEval is convenien...

DeepSpeed was secretly upcasting to FP32 — here's why your FSDP loss diverged — Inblix summary
Research

DeepSpeed was secretly upcasting to FP32 — here's why your FSDP loss diverged

Hugging Face Blog · Jun 13, 2024 · 2 min read

Running the same training pipeline with DeepSpeed and PyTorch FSDP should produce similar results. But when Hugging Fac...