Tag

#llms

Every item tagged llms, newest first.

22 items

We should be paid for using the internet.

A Reddit user argues that individuals should be compensated for contributing data to train LLMs, proposing payment based on the value contributed. The discussion reflects growing concerns about data ownership and fair compensation in the AI era. You may need to consider the implications of data usage and compensation in your own projects. This conversation is part of a broader debate on data rights and AI development.

Key takeaways

Users propose getting paid for data used in LLM training.
Payment could be based on value contributed.
Data ownership and compensation are growing concerns in AI.

rr/OpenAI#data-ownership #ai-ethics #llms

other15h

Quoting Charity Majors

Charity Majors argues that AI has made code generation effectively free and instant, turning the economics of code production upside down. This shift makes lines of code disposable and regenerable, rather than treasured and curated. Builders must adapt to this new reality. The change demands more engineering discipline, not less.

Key takeaways

Code generation is now effectively free and instant.
Lines of code are disposable and regenerable.
Builders need more engineering discipline with AI.

SSimon Willison#ai-assisted-programming #generative-ai #llms

research18h

Seeing Before Reasoning: Decoupling Perception and Reasoning for Shortcut-Resilient Multimodal On-Policy Self-Distillation

Researchers propose ViGOS, a visually grounded on-policy self-distillation framework for multimodal large language models (MLLMs). The method aims to prevent shortcuts in training where the model relies too heavily on text targets rather than image inputs. ViGOS guides the student model to use both visual and textual information effectively. This approach can improve the robustness of MLLMs in tasks that require multimodal understanding.

Key takeaways

ViGOS framework proposed for visually grounded on-policy self-distillation in MLLMs.
Method aims to prevent shortcuts relying on text targets over image inputs.
Improves robustness in multimodal tasks requiring both visual and textual understanding.

aarXiv#multimodal-learning #self-distillation #llms

research1d

RubricsTree: Scalable and Evolving Open-Ended Evaluation of Personal Health Agents across Health Memory and Medical Skills

Researchers propose RubricsTree, a framework for evaluating personal health agents powered by large language models. The framework addresses the challenge of scaling evaluation while maintaining clinical accuracy and consistency. RubricsTree aims to support the large-scale clinical deployment of these agents by providing a more efficient and reliable evaluation method.

Key takeaways

RubricsTree framework proposed for scalable evaluation of personal health agents.
Addresses bottleneck of physician annotation being costly and LLM evaluators being subjective.
Aims to support large-scale clinical deployment of LLM-empowered health agents.

aarXiv#healthcare #evaluation #personal-agents #llms

research1d

Evaluating Open-Source LLMs for Multi-Label ATT&CK Technique Classification on CTI Reports

Researchers evaluated open-source LLMs for multi-label ATT&CK technique classification on CTI reports. They found that LLMs can automate this complex task with high accuracy, reducing reliance on human effort. The study compared several open-source LLMs and provided insights into their performance on this specific task. You can apply these findings to improve CTI report classification in your own applications.

Key takeaways

Open-source LLMs achieve high accuracy in multi-label ATT&CK technique classification.
LLMs can automate complex CTI report analysis, reducing manual effort.
Study compared performance of multiple open-source LLMs on this task.

aarXiv#open-source #llms #cyber-threat-intelligence

tools1d

infiniflow/ragflow

infiniflow released ragflow, an open-source retrieval-augmented generation engine that combines rag and agent capabilities for llms. it provides a context layer to improve llm performance. builders can use ragflow to enhance their llm applications with advanced retrieval and generation features. ragflow is available on github for free.

Key takeaways

open-source rag engine with agent capabilities
fuses rag with agent tech for llms
available on github

research1d

hiyouga/LlamaFactory

The LlamaFactory repository provides a unified framework for efficient fine-tuning of over 100 large language models and vision-language models. This project was presented at ACL 2024. It offers a flexible and scalable solution for builders to adapt models to specific tasks. The repository is open-source and available on GitHub.

Key takeaways

Supports fine-tuning of 100+ LLMs and VLMs.
Presented at ACL 2024.
Open-source and available on GitHub.

modelsJun 9

llm 0.32a3

The llm 0.32a3 release was written almost entirely by the new Claude Fable 5 model. This marks a significant milestone in leveraging AI for content generation. The release demonstrates the capabilities of Claude Fable 5 in generating high-quality content. You can find more details in the author's write-up.

Key takeaways

Claude Fable 5 generated most of llm 0.32a3 release content.
New release showcases AI-driven content generation capabilities.
Author provides a detailed write-up of the process.

SSimon Willison#generative-ai #llms #claude-mythos

otherJun 1

Beyond LLMs: Why Scalable Enterprise AI Adoption Depends on Agent Logic

IBM Research and Hugging Face collaborated on agent logic as a key enabler of scalable enterprise AI adoption beyond LLMs. They argue that while LLMs are powerful, they often require additional logic to integrate with existing systems and processes. Builders should consider agent logic when designing AI systems for enterprise use cases.

Key takeaways

Agent logic is crucial for integrating LLMs with existing systems.
Scalable AI adoption depends on effective agent logic.
IBM Research and Hugging Face are collaborating on agent logic solutions.

HHugging Face Blog#enterprise-ai #agent-logic #llms

modelsApr 29

Granite 4.1 LLMs: How They’re Built

IBM released Granite 4.1, a series of open-weights LLMs. The models are trained on a mix of synthetic and human-generated data. IBM used a combination of automated and human evaluation to select the best model. You can access Granite 4.1 through Hugging Face.

Key takeaways

Trained on synthetic and human-generated data.
Uses automated and human evaluation.
Available on Hugging Face.

HHugging Face Blog#open-source #open-weights #llms

toolsSep 22

SyGra: The One-Stop Framework for Building Data for LLMs and SLMs

SyGra is a framework for generating data for large language models (LLMs) and small language models (SLMs). It aims to streamline data preparation and enable builders to create high-quality training data efficiently. SyGra is designed to work with various data sources and model architectures. You can use it to generate data for specific use cases or fine-tune existing models.

Key takeaways

SyGra streamlines data preparation for LLMs and SLMs.
Framework works with various data sources and model architectures.
Enables efficient creation of high-quality training data.

HHugging Face Blog#data-gen #llms #slms

researchSep 10

Jupyter Agents: training LLMs to reason with notebooks

Hugging Face released Jupyter Agents, a framework for training LLMs to interact with Jupyter notebooks. This enables models to reason over notebook contents and generate executable code. You can use Jupyter Agents to fine-tune models for domain-specific tasks, improving performance on tasks like data analysis and visualization.

Key takeaways

Jupyter Agents framework allows LLMs to interact with Jupyter notebooks.
Enables models to reason over notebook contents and generate code.
Fine-tuning with Jupyter Agents can improve model performance on domain-specific tasks.

HHugging Face Blog#jupyter #fine-tuning #domain-specific #llms

researchAug 12

TextQuests: How Good are LLMs at Text-Based Video Games?

Researchers evaluated leading LLMs on text-based video games, finding that even the best models struggle with long-term planning and common sense. The study used Hugging Face's Open LLM Leaderboard to assess model performance on TextQuests, a benchmark for text-based gaming. You can explore the results and leaderboard rankings on the Hugging Face blog. This assessment highlights areas where LLMs need improvement for practical applications.

Key takeaways

LLMs struggle with long-term planning and common sense in text-based games.
Study used Hugging Face's Open LLM Leaderboard and TextQuests benchmark.
Results show room for improvement in practical LLM applications.

HHugging Face Blog#llms #text-based-gaming #benchmarks

modelsApr 11

Visual Salamandra: Pushing the Boundaries of Multimodal Understanding

The Visual Salamandra 7B model is a new multimodal model that integrates text and image understanding. It is based on the LLaMA-2 architecture and has achieved state-of-the-art results on several benchmarks. The model is available on the Hugging Face platform for developers to use and build applications. You can leverage this model for tasks that require both text and image processing.

Key takeaways

Integrates text and image understanding
Based on LLaMA-2 architecture
Achieved state-of-the-art results on several benchmarks

HHugging Face Blog#multimodal #llms #hugging-face

otherApr 3

The NLP Course is becoming the LLM Course

The Hugging Face NLP Course is being renamed and updated to focus on large language models. The course will cover foundational concepts and practical applications of LLMs. You can expect new content on prompt engineering, model fine-tuning, and deployment. The updated course aims to equip you with skills to work with LLMs in real-world scenarios.

Key takeaways

The NLP Course is being renamed to the LLM Course.
The course will cover prompt engineering, fine-tuning, and deployment.
The update aims to address the growing demand for LLM skills.

HHugging Face Blog#education #llms #course

researchJan 23

Mastering Long Contexts in LLMs with KVPress

Researchers from NVIDIA and Hugging Face introduced KVPress, a method to improve long-context handling in large language models. KVPress uses a combination of techniques like sparse attention and compression to efficiently process longer sequences. This approach allows LLMs to handle up to 128K tokens, significantly expanding their context window. You can now explore KVPress in the Hugging Face Transformers library.

Key takeaways

KVPress enables LLMs to handle up to 128K tokens.
Uses sparse attention and compression for efficiency.
Available in Hugging Face Transformers library.

HHugging Face Blog#long-context #llms #transformers

modelsJan 9

CO₂ Emissions and Models Performance: Insights from the Open LLM Leaderboard

The Open LLM Leaderboard now displays CO₂ emissions for models, providing transparency on environmental impact. This allows you to compare models based on performance and emissions. The leaderboard tracks emissions from training and inference, giving insights into model efficiency. By considering both performance and environmental costs, you can make more informed decisions when selecting models for your applications.

Key takeaways

The Open LLM Leaderboard now shows CO₂ emissions for models.
Emissions data includes training and inference phases.
This helps compare models based on both performance and environmental impact.

HHugging Face Blog#open-source #sustainability #llms

researchNov 19

Judge Arena: Benchmarking LLMs as Evaluators

The Judge Arena benchmark evaluates LLMs as evaluators, comparing their ability to assess AI-generated text. The benchmark provides a framework for testing LLMs' evaluation capabilities, which is essential for developing reliable AI systems. You can use this benchmark to assess and compare the performance of different LLMs as evaluators. The benchmark's results can help you identify the strengths and weaknesses of various LLMs.

Key takeaways

Judge Arena benchmarks LLMs as evaluators of AI-generated text.
Provides a framework for testing LLMs' evaluation capabilities.
Helps identify strengths and weaknesses of LLMs as evaluators.

HHugging Face Blog#benchmarks #evaluators #llms

modelsJan 10

Make LLM Fine-tuning 2x faster with Unsloth and 🤗 TRL

Unsloth and Hugging Face's TRL library now enable 2x faster LLM fine-tuning. This integration allows builders to train models more efficiently. Faster fine-tuning means lower training costs and improved productivity. You can achieve these gains by leveraging the combined capabilities of Unsloth and TRL.

Key takeaways

Unsloth + TRL enables 2x faster LLM fine-tuning.
Integration reduces training costs and improves productivity.
Faster fine-tuning allows for more efficient model development.

HHugging Face Blog#fine-tuning #llms #hugging-face

modelsAug 8

Fine-tune Llama 2 with DPO

Hugging Face has released a tutorial on fine-tuning Llama 2 using Direct Preference Optimization (DPO). The tutorial covers implementing DPO with TRL, a popular open-source library for training and fine-tuning LLMs. You can use DPO to align model outputs with human preferences. This method provides an alternative to traditional reinforcement learning from human feedback (RLHF).

Key takeaways

DPO tutorial available for Llama 2 fine-tuning.
Uses TRL library for implementation.
DPO offers alternative to traditional RLHF.

HHugging Face Blog#fine-tuning #open-source #llms

toolsJul 24

Introducing Agents.js: Give tools to your LLMs using JavaScript

Agents.js is a new JavaScript library from Hugging Face that enables developers to give tools and abilities to large language models (LLMs). The library allows LLMs to interact with external systems and services using a simple and intuitive API. This enables builders to extend the capabilities of LLMs and create more sophisticated applications. By providing a straightforward way to integrate tools with LLMs, Agents.js aims to facilitate the development of more advanced and capable AI systems.

Key takeaways

Agents.js is a JavaScript library for giving tools to LLMs.
Enables LLMs to interact with external systems via a simple API.
Facilitates development of more advanced AI applications.

HHugging Face Blog#javascript #llms #tools

modelsJul 18

Llama 2 is here - get it on Hugging Face

Meta released Llama 2, a next-generation open-source LLM available for download on Hugging Face. The model targets research and commercial applications, offering improved performance and safety features. You can access Llama 2 through the Hugging Face model hub.

Key takeaways

Llama 2 is open-source and available on Hugging Face.
Targets both research and commercial use cases.
Includes improved safety features.

HHugging Face Blog#open-source #llms #hugging-face