Last week I've given a talk about building a React Code Agent for Transformers Agents! โค ๐ฅ๐ฒ๐๐ฐ๐: the agent iterates on previous actions in a ๐ฅ๐ฒflection-๐๐ฐ๐ion cycle โค ๐๐ผ๐ฑ๐ฒ: the agent writes actions as Code snippets, Read the slides
@aymericroucher
-
React LLM Agent Uses Memory to Solve Tasks
By
–
How can an agentic workflow use an LLM to solve tasks?
— m_ric (@AymericRoucher) 31 mai 2024
โก๏ธ I made my first ever ๐ฎ๐ข๐ฏ๐ช๐ฎ video to show that:
Watch below how a React LLM Agent solves a simple task, by leveraging its memory to iterate on previous actions! ๐บ๐ pic.twitter.com/a3S348qa0gHow can an agentic workflow use an LLM to solve tasks? I made my first ever ๐ฎ๐ข๐ฏ๐ช๐ฎ video to show that:
Watch below how a React LLM Agent solves a simple task, by leveraging its memory to iterate on previous actions! -
New ReactCodeAgents Available via Transformers Agents Package
By
–
Kind reminder that you can just 'pip install transformers[agents]' to try our shiny new ReactCodeAgents!
-
ICLR Paper Shows Code Improves LLM Agent Actions
By
–
ICLR Paper 'Executable Code Actions Elicit Better LLM Agents' shows when formulating Agent actions, ๐๐ผ๐ฑ๐ฒ ๐ถ๐ ๐ฏ๐ฒ๐๐๐ฒ๐ฟ than JSON or text for both ๐ฐ๐ผ๐ป๐ฐ๐ถ๐๐ฒ๐ป๐ฒ๐๐ ๐ฎ๐ป๐ฑ ๐ฝ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ฎ๐ป๐ฐ๐ฒ. This confirms our choice of using Code in Transformers Agents!
-
Guide to Structured JSON Generation for LLMs
By
–
New guide in our AI cookbook: ๐๐ฉ๐ง๐ช๐๐ฉ๐ช๐ง๐๐ ๐๐๐ฃ๐๐ง๐๐ฉ๐๐ค๐ฃ! This technique lets you force your LLM to generate its output as a JSON with specific keys: great for RAG or LLM-judge! Read it here: https://
huggingface.co/learn/cookbook
/structured_generation
โฆ Thank you @stevhliu for your advice! -
Hugging Face Transformers Update and AI Agent Guide
By
–
Accessible on Hugging Face Transformers `main` branch (v4.41.0 lands this week) If you've never played with Agents, the following guide gets you up to speed as to what's possible with them.
-
Agents 2.0 Framework Released with High Performance Benchmarks
By
–
Agents 2.0 is out, and it's already the best-performing framework using an open model ๐
— m_ric (@AymericRoucher) 13 mai 2024
โจ Simpler: prompt, tools, and attributes are accessible
๐งฉ Modular: use any LLM. Llama-3-70B-Instruct is ๐ฅ
๐ช Performant w/ React Agents
Top 1 of open models on GAIA, top 4 overall. pic.twitter.com/pzbVmuXysMAgents 2.0 is out, and it's already the best-performing framework using an open model Simpler: prompt, tools, and attributes are accessible Modular: use any LLM. Llama-3-70B-Instruct is Performant w/ React Agents Top 1 of open models on GAIA, top 4 overall.
-
Synthetic data from large models boost small models to top performance
By
–
Really useful! For a specific task, a synthetic dataset generated from a bigger model can be all you need to bring a smaller model all the way up to its bigger sibling's performance!
-
Small benchmark of 90 questions requiring only Search and Calculator
By
–
Note: this is only a small benchmark of 90 questions: 40 from gsm8k, 30 from HotpotQA, 20 from GAIA (cherrypicked to require only Search tool and Calculator).
So results may vary!
