deepset

company

Verified

https://deepset.ai

deepset_ai

deepset

Activity Feed

AI & ML interests

Semantic Search, Language models, Domain adaptation, Question Answering

Recent Activity

anakin87 updated a Space about 1 month ago

deepset/should-i-follow

anakin87 updated a Space about 1 month ago

deepset/autoquizzer

bogdankostic authored a paper 2 months ago

Same Meaning, Different Scores: Lexical and Syntactic Sensitivity in LLM Evaluation

View all activity

anakin87

posted an update 3 days ago

Post

3166

A small model that struggled against a random opponent now beats GPT-5-mini at tic-tac-toe

I took LiquidAI/LFM2-2.6B and trained it through play.

🧑‍🍳 Here's how:

1️⃣ Build a solid RL env with Verifiers (Prime Intellect)
2️⃣ Generate synthetic data: <200 games sampled from GPT-5-mini playing in the env
3️⃣ SFT warm-up to teach format
4️⃣ Group-based RL (CISPO) against opponents making 20-70% random moves
5️⃣ RL again with stronger opponents (0-25% random moves) + 1.25 temperature to push exploration and shake off suboptimal strategies

Done! Beats GPT-5-mini 🏆

---

🎮 Play against the model: anakin87/LFM2-2.6B-mr-tictactoe

🤗 Model: anakin87/LFM2-2.6B-mr-tictactoe

📚 Walkthrough/course: https://github.com/anakin87/llm-rl-environments-lil-course

🤗 Dataset and checkpoints: https://huggingface.co/collections/anakin87/lfm2-26b-mr-tic-tac-toe

anakin87

posted an update 4 days ago

Post

Local Gemma 4 agent 💎🕵️🗺️
drop in a mysterious map, get the location, live weather, and top spots to visit

I've been exploring what google/gemma-4-E4B-it can do in a local agentic setup and put together a 📓 𝙣𝙤𝙩𝙚𝙗𝙤𝙤𝙠 with Gemma + Haystack AI Framework covering 4 demos.

📓 https://t.ly/04Ty5

Another interesting one is the 𝗚𝗶𝘁𝗛𝘂𝗯 𝗔𝗴𝗲𝗻𝘁.

I initially tried to load all tools from the GitHub MCP server, quickly filling the context available on Colab -> unusable, forgetful agent ❌

Then I used the 𝗦𝗲𝗮𝗿𝗰𝗵𝗮𝗯𝗹𝗲 𝗧𝗼𝗼𝗹𝘀𝗲𝘁 🔎 🧰
It dynamically discovers the right tools from the GitHub MCP server on the fly, loading only what it actually needs for the task at hand, keeping context lean.

Now it actually works.

The notebook also contains
💎 Multimodal weather agent: the mystery map demo above
💎 Visual Question Answering from a paper
💎 RAG on Rock music

anakin87

posted an update 6 days ago

Post

10367

How LLM training with RL Environments works?

It all starts with 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗩𝗲𝗿𝗶𝗳𝗶𝗮𝗯𝗹𝗲 𝗥𝗲𝘄𝗮𝗿𝗱𝘀
- question asked
- model generates reasoning + answer
- answer checked against ground truth
- reward drives RL training

In this setup, the environment is simple: fixed questions and answers, rollout logic, reward(s)

Consider a more complex tic-tac-toe env ❌⭕
It adds:
- dynamic game generation/handling
- tunable opponent skill
- multi-turn interactions

(envs can also include tools)

---

What happens at training?

We use 𝗚𝗿𝗼𝘂𝗽 𝗥𝗲𝗹𝗮𝘁𝗶𝘃𝗲 𝗣𝗼𝗹𝗶𝗰𝘆 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗮𝘁𝗶𝗼𝗻 with a tic-tac-toe env

No critic model needed, the group is the baseline
Simpler than PPO

1️⃣ Rollout generation: from the same board, model plays N games via sampling
2️⃣ Each game scored with deterministic rewards (win, format, ...)
3️⃣ Mean score computed across the group
4️⃣ Each rollout's advantage = its score minus the group mean
5️⃣ Model updated to favor trajectories above baseline

🔁 Repeat

For a deep dive, check out
🌱 https://github.com/anakin87/llm-rl-environments-lil-course
a free hands-on course on RL environments for LLMs

2 replies

anakin87

posted an update 10 days ago

Post

1613

Your RL environment is an SFT data factory 🏭

In LLM post-training it's common to do Supervised Fine-Tuning warm-up before Reinforcement Learning.

When teaching a new task, RL needs some signal to amplify and SFT builds a good initial basis, for example by teaching format.

If you've built an RL env, generating SFT synthetic data is basically free.

An env already has: task data, rollout logic, rewards.

1️⃣ pick a strong model
2️⃣ run it through the env
3️⃣ filter rollouts by reward

works out of the box with Verifiers (Prime Intellect) and Atropos (Nous Research)

🧑‍💻 Example: https://github.com/anakin87/llm-rl-environments-lil-course/blob/main/chapters/05.md

anakin87

posted an update 15 days ago

Post

4144

🌀 Let LLMs wander - Engineering RL Environments

Reinforcement Learning Environments are little worlds
where models can act, get rewards, and learn.

I've been exploring how to design them, figuring out what works and what doesn't.

If you want to learn how to build them, I recorded a practical intro video.

You'll also see how to turn Liquid AI LFM2-2.6B into a Tic-tac-toe master 🙂

🎥 Engineering RL Environments video: https://www.youtube.com/watch?v=71V3fTaUp2Q

---

🌱 LLM RL Environments Lil Course: https://github.com/anakin87/llm-rl-environments-lil-course

🤗🕹️ Play against the trained model: anakin87/LFM2-2.6B-mr-tictactoe

📚 HF collection (datasets + models): https://huggingface.co/collections/anakin87/lfm2-26b-mr-tic-tac-toe

anakin87

posted an update 18 days ago

Post

3296

📣 I just published a free course on Reinforcement Learning Environments for Language Models!

📌 COURSE: https://github.com/anakin87/llm-rl-environments-lil-course

Over the past year, we've seen a shift in LLM Post-Training.
Previously, Supervised Fine-Tuning was the most important part: making models imitate curated Question-Answer pairs.

Now we also have Reinforcement Learning with Verifiable Rewards. With techniques like GRPO, models can learn through trial and error in dynamic environments. They can climb to new heights without relying on expensively prepared data.

But what actually are these environments in practice❓ And how do you build them effectively❓

Fascinated by these concepts, I spent time exploring this space through experiments, post-training Small Language Models.
I've packaged everything I learned into this short course.

What you'll learn

🔹 Agents, Environments, and LLMs: how to map Reinforcement Learning concepts to the LLM domain
🔹 How to use Verifiers (open-source library by Prime Intellect) to build RL environments as software artifacts
🔹 Common patterns: How to build single-turn, multi-turn, and tool-use environments

🔹 Hands-on: turn a small language model (LFM2-2.6B by LiquidAI) into a Tic Tac Toe master
🔸 Build the game Environment
🔸 Use it to generate synthetic data for SFT warm-up
🔸 Group-based Reinforcement Learning

If you're interested in building "little worlds" where LLMs can learn, this course is for you.

---

🤗🕹️ Play against the trained model: anakin87/LFM2-2.6B-mr-tictactoe

📚 HF collection (datasets + models): https://huggingface.co/collections/anakin87/lfm2-26b-mr-tic-tac-toe

1 reply

anakin87

updated 2 Spaces about 1 month ago

Should I follow?

🦄

Find recent Bluesky posts for any user

AutoQuizzer

🧑

225

Generate a quiz, play or let 🦙 LLM play

bogdankostic

authored a paper 2 months ago

Same Meaning, Different Scores: Lexical and Syntactic Sensitivity in LLM Evaluation

Paper • 2602.17316 • Published Feb 19 • 1

bilgeyucel

updated a Space 2 months ago

README

🐠

anakin87

posted an update 5 months ago

Post

394

💭 Do thinking traces make Language Models learn better? Curious what others think

𝗦𝗰𝗲𝗻𝗮𝗿𝗶𝗼
You take an instruction-following LM.
You want to train it with a GRPO-style RL algorithm on a task like Tic Tac Toe.
Rewards are outcome-based, applied only at the end of each episode: win/loss/draw, format adherence...

During training, the model could just output answers, but a common choice is to make it also output thinking traces.

𝗧𝗵𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻
Does forcing the model to produce thinking traces during training actually improve learning❓

💬 I'd like to hear your thoughts. Share ideas and links to relevant papers and resources.

From what I've understood so far, the answer seems to be 𝘆𝗲𝘀.

1️⃣ If you force the model to think during training, it becomes a model that thinks at inference time. It naturally allocates more budget (tokens) to a problem, which tends to improve performance.

2️⃣ While the model's "reasoning" already exists in its activation space, using explicit thinking traces as a scratchpad allows training to steer and shape that reasoning.

3️⃣ As the model produces more traces during training, the RL algorithm can progressively give higher rewards to the reasoning patterns that lead to better outcomes.

anakin87

posted an update 5 months ago

Post

472

I made a visualization based on the Prime Intellect INTELLECT-3 technical report.

Wild to see how far they pushed GLM-4.5-Air-Base with SFT + RL.
SOTA for its size and competitive with models 3x larger.

All open.

Congrats on the release!

Model: PrimeIntellect/INTELLECT-3
Technical report: https://storage.googleapis.com/intellect-3-paper/INTELLECT_3_Technical_Report.pdf
Chat: https://chat.primeintellect.ai/

anakin87

posted an update 5 months ago

Post

2898

LLMs can leak their post-training data (RL included) 💧

New interesting paper on this topic from Google DeepMind: Extracting alignment data in open models (2510.18554)

It's known that Language Models memorize data that can be extracted via prompting.

In this paper, the authors investigate this aspect:
- using open models, where prompting can be fully customized by the user, including special tokens.
- focusing on open-source models like Olmo, where full training data is available.

📤 How do they extract data?

During post-training (like SFT), new tokens such as <|user|> are introduced.

The authors hypothesize prompting the model with these tokens can make it output its alignment data (remember Magpie?).

For example, for SFT, their extraction prompt is <|endoftext|><|user|>.

📏 Evaluating memorization

The authors compare each sampled example with the original data using vector search with embedding similarity.

They find that many outputs are semantically very similar to the original data, even if the exact words differ.

Traditional string-matching algorithms underestimate memorization by 10x.

🔁 What about RL?

Surprisingly, the same technique works to extract data from Reinforcement Learning (PPO/GRPO) phases.

This is counter-intuitive because the RL objective is not designed to increase sequence likelihoods (unlike SFT).

Practical limitation: in this case, extraction relies on using the initial part of the training prompt, which is not generally public.

📈 Are the extracted data effective for post-training?

Both in SFT and RL, the extracted data can be used to fine-tune models to similar performance to the originals.

The authors suggest that model distillation, where a stronger model is used to drive the training of a weaker one, may be a form of indirect training on the original dataset.

anakin87

posted an update 8 months ago

Post

506

Your Language Model needs better (open) environments to learn 🌀

📝 https://huggingface.co/blog/anakin87/environments-hub

RL environments help LLMs practice, reason, and improve.
I explored the Environments Hub and wrote a walkthrough showing how to train and evaluate models using these open environments.

1️⃣ 𝗪𝗵𝘆 𝗥𝗟 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀

DeepSeek-R1 made clear that Reinforcement Learning can be used to incentivize reasoning in LLMs.
In GRPO, the model generates multiple answers and learns to prefer the better ones from rewards.

2️⃣ 𝗪𝗵𝗮𝘁 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝘀 𝗮𝗿𝗲
In classic RL, the environment is the world where the Agent lives, interacts, and get rewards to learn.

We can also think of them as software packages, containing data, harness and scoring rules - for the model
to learn and be evaluated.

Nowadays, the Agent is not just the LLM. It can use tools, from a weather API to a terminal.

This makes environments for training and evaluation more complex and critical.

3️⃣ 𝐓𝐡𝐞 𝐨𝐩𝐞𝐧 𝐜𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞

Big labs are advancing, but open models and the community still face a fragmented ecosystem.
We risk becoming users of systems built with tools we can't access or fully understand.

4️⃣ 𝐄𝐧𝐯𝐢𝐫𝐨𝐧𝐦𝐞𝐧𝐭𝐬 𝐇𝐮𝐛
That's why, I was excited when Prime Intellect released the Environments Hub.

It's a place where people share RL environments: tasks you can use to train LLMs with RL (GRPO-style) or evaluate Agents.
Plus, the Verifiers library (@willcb ) standardizes the creation of RL environments and evaluations.
They can help to keep science and experimentation open. 🔬

I explored the Hub and wrote a hands-on walkthrough 📝
- RL + LLMs basics
- Environments Hub navigation
- Evaluating models/Agents
- GRPO Training a tiny model on an alphabetical sort task

Take a look!

📝 https://huggingface.co/blog/anakin87/environments-hub

anakin87

posted an update 8 months ago

Post

4747

Want to quickly try Gemma 3 270m? 💎💬

I made a simple Space to do that: anakin87/gemma-3-270m-it

⚡ Fast: Flash Attention, Zero GPU
⚙️ Configurable

anakin87

posted an update 9 months ago

Post

405

🕵️🌐 Building Browser Agents - notebook

No API? No problem.
Browser Agents can use websites like you do: click, type, wait, read.

📓 Step-by-step notebook: https://colab.research.google.com/github/deepset-ai/haystack-cookbook/blob/main/notebooks/browser_agents.ipynb

🎥 In the video, the Agent:
- Goes to Hugging Face Spaces
- Finds black-forest-labs/FLUX.1-schnell
- Expands a short prompt ("my holiday on Lake Como") into a detailed image generation prompt
- Waits for the image
- Returns the image URL

## What else can it do?
Great for information gathering and summarization

🗞️🗞️ Compare news websites and create a table of shared stories with links
▶️ Find content creator social profiles from YouTube videos
🛍️ Find a product's price range on Amazon
🚂 🚌 Gather public transportation travel options

## How is it built?
🏗️ Haystack → Agent execution logic
🧠 Google Gemini 2.5 Flash → Good and fast LLM with a generous free tier
🛠️ Playwright MCP server → Browser automation tools: navigate, click, type, wait...

Even without vision capabilities, this setup can get quite far.

## Next steps
- Try a local open model
- Move from notebook to real deployment
- Incorporate vision

And you? Have you built something similar? What's in your stack?

anakin87

posted an update 9 months ago

Post

1098

Haystack can now see 👀

The latest release of the Haystack OSS LLM framework adds a long-requested feature: image support!

📓 Notebooks below

This isn't just about passing images to an LLM. We built several features to enable practical multimodal use cases.

What's new?
🧠 Support for multiple LLM providers: OpenAI, Amazon Bedrock, Google Gemini, Mistral, NVIDIA, OpenRouter, Ollama and more (support for Hugging Face API coming 🔜)
🎛️ Prompt template language to handle structured inputs, including images
📄 PDF and image converters
🔍 Image embedders using CLIP-like models
🧾 LLM-based extractor to pull text from images
🧩 Components to build multimodal RAG pipelines and Agents

I had the chance of leading this effort with @sjrhuschlee (great collab).

📓 Below you can find two notebooks to explore the new features:
󠁯•󠁏󠁏 Introduction to Multimodal Text Generation https://haystack.deepset.ai/cookbook/multimodal_intro
󠁯•󠁏󠁏 Creating Vision+Text RAG Pipelines https://haystack.deepset.ai/tutorials/46_multimodal_rag

(🖼️ image by @bilgeyucel )

anakin87

posted an update 10 months ago

Post

458

🛡️ AI Guardrails with Open Language Models - Tutorial

📓 https://haystack.deepset.ai/cookbook/safety_moderation_open_lms

How do you ensure your AI application is safe from harmful or inappropriate user inputs?

This is a core requirement for real-world AI deployments. Luckily, several open Language Models are built specifically for safety moderation.

I've been exploring them and put together a hands-on tutorial using the Haystack framework to build your own AI guardrails.

In the notebook, you'll learn how to use and customize:
🔹 Meta Llama Guard (via Hugging Face API)
🔹 IBM Granite Guardian (via Ollama), which can also evaluate RAG specific risk dimensions
🔹 Google ShieldGemma (via Ollama)
🔹 Nvidia NemoGuard models family, including a model for topic control

You'll also see how to integrate content moderation into a 🔎 RAG pipeline.

anakin87

posted an update 10 months ago

Post

1234

🧰 Free up space on the Hub with super_squash_history 🧹

As you may know, Hugging Face Hub has storage limits on private repos (100 GB for free users, 1 TB for PROs).

This weekend I did some cleanup on my private repos
I went 1.58 TB down to 1 GB. 😅

Besides deleting old, unused models, the main tool I used was a lesser-known command:
super_squash_history.

When you train a model, you often push multiple checkpoints to the Hub.
Each checkpoint = a commit.
A 2.6B model in BF16 is ~5 GB.
So 10 checkpoints = 50 GB. That adds up fast.

While full commit history can be useful for rollbacks, it's often unnecessary for older experiments where only the final model matters.

In these cases, you can use super_squash_history: it reduces your entire repo history to a single commit.

https://huggingface.co/docs/huggingface_hub/main/en/package_reference/hf_api#huggingface_hub.HfApi.super_squash_history

⚠️ super_squash_history is a non-revertible operation. Once squashed, the commit history cannot be retrieved.

Hope this is useful to others.