Stefano Fiorucci

🌱 LLM RL Environments free course: | AI/SW Engineer 🧑‍🚀 | Building @haystack_ai OSS LLM framework | Tinkering with Post-Training, RL, SLM

[Stefano Fiorucci @theanakin87]

the LLM RL Environments Lil Course is live!

a little course on reinforcement learning environments for evaluating and training language models

with @PrimeIntellect Verifiers + @liquidai models

go check it out: https://t.co/slVSmXviki

[Stefano Fiorucci @theanakin87]

you can just build RL Environments

COMING SOON

[Stefano Fiorucci @theanakin87]

@lele_venturi Ahah 💙

[Stefano Fiorucci @theanakin87]

Emma 5 è tornata! 🙂

https://t.co/mI0LUPU3dJ

Si tratta di piccolo modello italiano noto più per i meme che per le sue capacità.

Io e Claude ci siamo divertiti a rimetterlo online.

Gira sulla GPU di Hugging Face. Se il sito vi blocca, basta loggarsi con un account gratuito.

[Stefano Fiorucci @theanakin87]

@Icarussiano Ci sta... Cmq mi sono divertito!

[Stefano Fiorucci @theanakin87]

@Sofia_Phobia Se avete buone possibilità di avere un reddito più alto in futuro, può avere senso iniziare a investire somme piccole per imparare/misurarsi con gli strumenti... Che ne pensi?

[Stefano Fiorucci @theanakin87]

have you tried the Google Colab CLI?

a programmatic way to access Colab from the shell for humans and agents

I finally tried it.

I had outdated notebooks to refresh and wanted to use Claude Code, but also needed GPU access in the actual Colab environment.

Works well!

[Stefano Fiorucci @theanakin87]

@helloiamleonie @liquidai @maximelabonne @paulabartabajo_ Oooh congrats!

[Stefano Fiorucci @theanakin87]

Pycon Italy starts soon and I'll be there!

On Saturday, I'll give a talk on Reinforcement Learning Environments 🌍:
little worlds where language models can act, get rewards, and learn.

Curious about model training beyond SFT? Join this seesion.

See you in Bologna...

@pyconit

[Stefano Fiorucci @theanakin87]

@itsafiz thx! vLLM + Haystack = 💕
😁

[Stefano Fiorucci @theanakin87]

interesting line of research

early stop RLVR and extrapolate future weights updates

curious whether this holds on other model families, given Qwen2.5 suspect contamination on math benchmarks (https://t.co/U6ze5yN9Bf)

[Zhepei Wei @weizhepei]

😢RLVR is powerful but expensive
🤯Imagine using <20% RLVR training while achieving 100% performance?

Sounds surprising? We show that minimal RLVR training is enough to know where training is going, and predict future ckpts at no training cost!

📃https://t.co/fGODWWIjR1
🧵[1/n]

[Stefano Fiorucci @theanakin87]

@itsafiz @vllm_project @Haystack_AI Oh great! If something doesn't work as expected, let me know...

[Stefano Fiorucci @theanakin87]

vLLM x Haystack 🤝

@vllm_project is the standard for serving LLMs on GPU
now @Haystack_AI ships a native integration

⚡ serve models fast: generative LLMs, embeddings and rerankers
🏗️ build AI apps and agents on top

🐍 pip install vllm-haystack

notebook + resources below 👇

[Stefano Fiorucci @theanakin87]

Notebook: https://t.co/3ehWpZ8QT8
Integration page: https://t.co/ck9xnxkSQt

[Bilge @bilgeycl]

say hi to the @Haystack_AI open source team 👋🏼

We just wrapped up a week in Berlin at @deepset_ai hq, can’t wait to share more soon!

[Stefano Fiorucci @theanakin87]

@27upon2 @chakra_ai @GovViz oooh congrats!!!

[Stefano Fiorucci @theanakin87]

@DaKulchur @googlegemma @Haystack_AI @vladblagoje Haystack here helps with the following:
- ready-made agent abstraction
- searchable toolset (loads tools when needed)
- provider-agnostic (in the notebook I use Ollama, easy to switch to llama.cpp, vLLM...)

[Stefano Fiorucci @theanakin87]

local Gemma 4 agent: drop in a map, get the location, live weather, and top spots to visit

put together a notebook with @googlegemma + @Haystack_AI covering the above and

1. GitHub Agent: discovers the right tools from MCP on the fly, keeping context lean (h/t @vladblagoje)
...

[Stefano Fiorucci @theanakin87]

@CoreyGallon @aiDotEngineer Thank you!

[Corey J. Gallon @CoreyGallon]

A 2.6B parameter model trained with reinforcement learning beat GPT-5 mini at tic-tac-toe.

Not by using better data. By using a better environment.

@theanakin87 at @aiDotEngineer Europe on why the next leap for open models isn't datasets -- it's gyms.

[Sinatras @ ICML @myainotez]

This moment feels special, we will see glimpse of what can happen if agents utilized in an adversarial setting for autoresearch. You dont need to know anything about kernels you agent already knows it, just give them the key via flywheel and let them go wild

[Stefano Fiorucci @theanakin87]

@IALatino @googlegemma @Haystack_AI @vladblagoje Gemma 4 E4B: runs great on Colab

[Stefano Fiorucci @theanakin87]

@sodakeyEatsMush might be wrong but I'd try increasing batch_size (especially) and num_generations

different application, but you can find sth useful on my lil course: https://t.co/slVSmXviki