- Blablador-news - lists.fz-juelich.de

QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning
by Strube, Alexandre 05 Jun '25

05 Jun '25

From the huggingface repo: QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning We release 🤗 QwenLong-L1-32B, which is the first long-context LRM trained with reinforcement learning for long-context reasoning. Experiments on seven long-context DocQA benchmarks demonstrate that QwenLong-L1-32B outperforms flagship LRMs like OpenAI-o3-mini and Qwen3-235B-A22B, achieving performance on par with Claude-3.7-Sonnet-Thinking, demonstrating leading performance among state-of-the-art LRMs. We have it on Blablador. I don’t have yet a perfect system prompt for it, so you might see some funny tokens there, like <|im_end|> and so on. In any case, it would be nice to see if it’s useful for you, and then I would keep it running. Please let me know! Let’s bark! Alex

1 0

Hi
by Na, Sehyang 27 May '25

27 May '25

1 0

Helmholtz services downtime on 2025-05-22 16:00 CEST, for approx. 1 hour.
by Strube, Alexandre 20 May '25

20 May '25

This is coming from the HIFIS team: > Dear Helmholtz Cloud users: > Helmholtz Codebase [1] and Mattermost [2] will undergo planned maintenance and will not be accessible from 2025-05-22 16:00 CEST, for approx. 1 hour. > In preparation for the GitLab major upgrade 18, a planned maintenance of the database is taking place. Access to both services is not possible during this time. > With regards, > the HIFIS team In practice: Blablador’s API server won’t be able to authenticate the token, so it will fail. I’m sorry for the disturbance. Best, Alex

1 0

Llama3-405 is out for tomorrow
by Strube, Alexandre 14 May '25

14 May '25

Hello Blabladores, as the West AI machine that we run Llama3-405 on is on testing tomorrow, you probably noticed that the model is out, and it will be for the whole day tomorrow. Sorry for the inconvenience. Alex

1 0

Attempting to phase out llama3.3 for alias-large in favor of Qwen3-30B-A3B
by Strube, Alexandre 05 May '25

05 May '25

Hello Blabladores! In keeping with the latest developments in LLMs, I decided to try Qwen-3 30B - A3B. This is a much smaller model than Llama 3.3 - 70B, as well as as the previous model running as alias-large, DeepSeek-R1-70B, which was not so good and very unstable. This is a Mixture of Experts (MoE) model, with 30 billion parameters, where each expert has 3 billion parameters. This means a number of things: - Given the model is much smaller, we can have a much bigger context size, as there is more GPU memory available. The consquence is that the model is more useful with bigger workloads and agents. While I had to keep Llama 3.3-70 limited to a maximum context size of 8192 tokens, I can let Qwen-3 run at its maximum size, which is 128k tokens. - The mixture of experts model architecture means that, when activated, you are performing inference only on one expert at a time, and this is of 3 billion parameters - meaning it is MUCH faster, which makes it again more useful. Qwen3 is a much more modern model than the Llama3.3 - Five months in this field can make a huge difference. Besides, Qwen3 uses the Apache-2.0 license, and I am much more partial to it than to the arcane Meta Llama License which should not even exist. Mind you that while this is a reasoning model, I had to remove the reasoning part on the chat template - This is because the thinking process was causing the web ui to fail intermittently. When I get a fix on that, I will remove the setting so you can see the reasoning process. This affects both the Web UI and the API access. The chat template have a sep="<|im_end|>/nothink", so the model never reasons. If you have any feedback on the workings of the model, or if you think the quality has suffered, please let me know right away! In the worst case I would revert to Llama3.3. But I hope it won’t be needed! Let’s bark! Alex

1 0

Web UI Fixed!
by Strube, Alexandre 15 Apr '25

15 Apr '25

Turns out that this was a confusing one! A while ago, I proposed changes to Gradio, the library which powers Blablador’s web ui, so we could generate a watermark when copying text from it. This is to comply with the EU AI Act. They took the change in. So I updated the library on the server. All good. At the same time, I removed the message about the 10-year old birthday some of you might have seen around. But then, things got weird: It was suddenly happening that I had to write something first, then write again, so the system would respond! As the two things were done roughly at the same time, I blamed Gradio - it wouldn’t have been the first time. To make it even weirder, after wasting a night on this, I realized that SOME models worked. After an even longer time, my investigation got weirder: models which didn’t work would work if I changed their names to something else. At this point I was already running the server line by line on a debugger, and then it hit me: the system prompt of the birthday: # A template with a one-shot conversation example register_conv_template( Conversation( name="one_shot", system_message="A chat between a curious human and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the human's questions.", roles=("Human", "Assistant"), messages=( ( "Human", "Got any creative ideas for a 10 year old’s birthday?", ), ( "Assistant", """Of course! Here are some creative ideas for a 10-year-old's birthday party: 1. Treasure Hunt: Organize a treasure hunt in your backyard or nearby park. Create clues and riddles for the kids to solve, leading them to hidden treasures and surprises. 2. Science Party: Plan a science-themed party where kids can engage in fun and interactive experiments. You can set up different stations with activities like making slime, erupting volcanoes, or creating simple chemical reactions. 3. Outdoor Movie Night: Set up a backyard movie night with a projector and a large screen or white sheet. Create a cozy seating area with blankets and pillows, and serve popcorn and snacks while the kids enjoy a favorite movie under the stars. 4. DIY Crafts Party: Arrange a craft party where kids can unleash their creativity. Provide a variety of craft supplies like beads, paints, and fabrics, and let them create their own unique masterpieces to take home as party favors. 5. Sports Olympics: Host a mini Olympics event with various sports and games. Set up different stations for activities like sack races, relay races, basketball shooting, and obstacle courses. Give out medals or certificates to the participants. 6. Cooking Party: Have a cooking-themed party where the kids can prepare their own mini pizzas, cupcakes, or cookies. Provide toppings, frosting, and decorating supplies, and let them get hands-on in the kitchen. 7. Superhero Training Camp: Create a superhero-themed party where the kids can engage in fun training activities. Set up an obstacle course, have them design their own superhero capes or masks, and organize superhero-themed games and challenges. 8. Outdoor Adventure: Plan an outdoor adventure party at a local park or nature reserve. Arrange activities like hiking, nature scavenger hunts, or a picnic with games. Encourage exploration and appreciation for the outdoors. Remember to tailor the activities to the birthday child's interests and preferences. Have a great celebration!""", ), ), offset=2, sep_style=SeparatorStyle.ADD_COLON_SINGLE, sep="\n### ", stop_str="###", ) I could just delete this prompt, right? WRONG Look at another one: # A template similar to the "one_shot" template above but remove the example. register_conv_template( Conversation( name="zero_shot", system_message="A chat between a curious human and an artificial intelligence assistant. " "The assistant gives helpful, detailed, and polite answers to the human's questions.", roles=("Human", "Assistant"), sep_style=SeparatorStyle.ADD_COLON_SINGLE, sep="\n### ", stop_str="###", ) ) This is the culprit: offset=2, This skips showing the two first messages on the dialogue, since they are system messages and should be hidden. Which meant that the message I sent and the first reply would never show up. Easy, right? Then I realized I was saving the files on my test instance and had my browser open on the production instance, which was not working because of course it wasn’t! I wasn’t changing anything in there. So, that’s it. A simple change that made a hell of a weekend and made me lost some days of work. But now you won’t hear about the birthday anymore :-D Let’s bark! Dr. Alexandre Strube a.strube(a)fz-juelich.de Helmholtz AI Jülich Supercomputing Centre Forschungszentrum Juelich GmbH 52425 Jülich, Germany Phone: +49 2461 61-3866 JSC is the coordinator of the John von Neumann Institute for Computing (NIC) and member of the Gauss Centre for Supercomputing (GCS)

1 0

Problem on the UI
by Strube, Alexandre 14 Apr '25

14 Apr '25

There’s a little problem with the Web UI, in which the first message to the system doesn’t appear on the screen. I’m on it! Dr. Alexandre Strube a.strube(a)fz-juelich.de Helmholtz AI Jülich Supercomputing Centre Forschungszentrum Juelich GmbH 52425 Jülich, Germany Phone: +49 2461 61-3866 JSC is the coordinator of the John von Neumann Institute for Computing (NIC) and member of the Gauss Centre for Supercomputing (GCS)

1 0

DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level
by Strube, Alexandre 09 Apr '25

09 Apr '25

This model, released last night, is quite something: It’s a qwen-14 distilled by deepseek r1, and then fine-tuned to code by Together AI. Together AI claims they match o3-mini-2025-01-031 (Low) and o1-2024-12-17. The good thing is that they opened the source AND data so people can see what they did there. And obviously, this is now the default code model on Blablador! It’s a reasoning model, so you will hear echoes of the system prompt (birthday party) in there. It also have a 64k context on our hardware, which is nice! Here is the report: https://www.together.ai/blog/deepcoder Let’s keep barking! Alex

1 0

And Llama4 is... out.
by Strube, Alexandre 07 Apr '25

07 Apr '25

If people are following the news, turns out that Llama-4 came with a thud. It’s not a good model and it seems like they cheated on benchmarks. One of the examples is this: Plus, their license doesn’t allow for multi-modals models to run in Europe - And even though I don’t provide means to have them using this ability, I prefer to leave them out for now. We have many other excellent models to offer! Alex

1 0