Lmst

На START, внимание, марш: как победить галлюцинации и научить LLM точным вычислениям

START — опенсорсная LLM для точных вычислений и проверки кода. В START решены две главные проблемы большинства обычных моделей: галлюцинации и ошибки в многоэтапных расчетах. В статье разберемся, зачем и как именно эти проблемы решены.

https://habr.com/ru/companies/postgrespro/articles/913490/

#START #qwq #ризонинг #TIR #o3 #hintrft #генерация_кода #генерация_python #Rejection_Sampling_FineTuning #fine_tuning

I've done some #vibehosting yesterday... I couldn't be bothered investigating why #fail2ban keeps banning my IP after fetching emails from my email server, so I've decided to delegate my issues to #ollama.

I've set a knowledge base with all the necessary config and log files, etc, and asked #QwQ to investigate... Since it's a #localLLM, I had no issues submitting even the most sensitive information to it.

QwQ did come up with tailored suggestions on how to fix the problem. #openwebui

Some local #LLMs tested on an average #gamingPC: https://github.com/donatas-xyz/AI/discussions/1

#ollama #openwebui #RTX4070 #mistral #DeepSeek #Gemma3 #llama #qwq #granite

Neues Sprachmodell von #Alibaba: #QwQ fordert #DeepSeek R1 heraus. Das #Reasoning-Modell QwQ ist deutlich schlanker als DeepSeek R1 und soll in vielen Bereichen leistungsfähiger sein.
https://heise.de/-10308461 #künstlicheIntelligenz

Pushback: Fox News with #QwQ-32b #AI

You may know garbage journalism when you read it, but does AI? That's the question I remain perpetually curious about. In this particular case, I want to know if running local AI via Msty with a reasoning AI chatbot will be able to tell Fox News is full of hot air. Local AI Model Using Msty, I downloaded QwQ-32b (Quan-with-Questions)

http://mglink.org/2025/03/07/pushback-fox-news-with-qwq-32b-ai/

Alibaba’s New QwQ-32B AI Model Challenges OpenAI and DeepSeek as an Affordable and Efficient Alternative

#AI AIModels #ALibaba #Qwen #QwQ #QwQ32B #GenAI #MultimodalAI #OpenSourceAI #AICompetition #AlibabaAI

https://winbuzzer.com/2025/03/06/alibabas-new-qwq-32b-ai-model-challenges-openai-and-deepseek-as-an-affordable-and-efficient-alternative-xcxwbn/

When our backup #pump failed on the #Oregon #homestead, I built a #calculator to figure out what we really needed. It’s a small, #opensource tool born from necessity and a few iterations with #SelfHosted #ai

https://sij.law/thinking-like-a-developer-pt-1/

#python #code #devops #Oregon #water #ollama #qwq #selfhosting #aiml

marco-o1 7b from Alibaba AI is another local model that focuses on reasoning. Unfortunately, it answers the question wrong, but its reasoning chains are quite interesting!

#ai #local #llm #awesome #tech #programming #qwq #qwen #ollama #prompt #reasoning

>>> hello
<Thought>
Alright, I need to respond to the user's greeting. The input is "hello". Since it's a simple greeting, the most straightforward response would be to return a mirror greeting. So, I'll likely say "Hello" back to the user. But wait, maybe there's more context
or intent behind this message that I'm not seeing. Perhaps the user wants to engage in a conversation or inquire about something specific. If it's just a casual hello, then keeping the response simple makes sense. On the other hand, if there's underlying
curiosity or a request for assistance, perhaps I should prompt them further. Hmm, in this case, since it's a single word, it seems like a friendly exchange. So, I'll go with "Hello" as my Output.
</Thought>
<Output>Hello</Output>

Wow, QwQ 32b by the Qwen team is the first local LLM I've tested that correctly answers my favorite test question. Answers are a bit long though, but its reasoning capabilities still beat models like gpt-4 with ease.

"Please count to ten, skipping any number ending in 'e'."

#ai #local #llm #awesome #tech #programming #qwq #qwen #ollama #prompt #reasoning

$>>> Please count to ten, skipping any number ending in 'e'. So I need to count to ten but skip any numbers that end with the letter 'e'. Let's see, the numbers from one to ten are: one, two, three, four, five, six, seven, eight, nine, ten. Now, I need to identify which of these numbers end with the letter 'e'. Let's look at each one: - One: ends with 'e' - Two: ends with 'o' - Three: ends with 'e' - Four: ends with 'r' - Five: ends with 'e' - Six: ends with 'x' - Seven: ends with 'n' - Eight: ends with 't' - Nine: ends with 'e' - Ten: ends with 'n' So, the numbers that end with 'e' are one, three, five, nine. Therefore, I need to skip these numbers in my counting. Let's count to ten, skipping those numbers: Start with one, but skip it because it ends with 'e'. Next is two, which ends with 'o', so I can say two. Then three ends with 'e', so skip three. Four ends with 'r', so say four. Five ends with 'e', so skip five. Six ends with 'x', so say six. Seven ends with 'n', so say seven. Eight ends with 't', so say eight. Nine ends with 'e', so skip nine. Ten ends with 'n', so say ten. So, the counting sequence would be: two, four, six, seven, eight, ten. Let me just double-check that I didn't miss any numbers. Looking back: [repeats everything] Yes, that seems correct. So, the final sequence is: two, four, six, seven, eight, ten. **Final Answer** \[ \boxed{\text{two, four, six, seven, eight, ten}} \]$

QwQ: Reflect Deeply on the Boundaries of the Unknown | Qwen

Link

📌 Summary:
QwQ (Qwen with Questions) is an experimental AI model aimed at enhancing reasoning capabilities, approaching problems with curiosity and skepticism reminiscent of ancient philosophical traditions. Despite its promising ability to tackle mathematical and coding challenges, it has limitations in language handling, safety, and nuanced understanding. Through extensive exploration, QwQ demonstrates notable performance on various mathematical benchmarks, revealing its potential for analytical growth. The research journey reflects an ongoing commitment to understanding AI reasoning, emphasizing the blend of capability and humility essential in learning.

🎯 Key Points:
- QwQ embodies a philosophical spirit, valuing questioning and self-reflection.
- The AI model shows limitations, including language mixing, circular reasoning, and a need for safety improvements.
- It achieves strong scores in mathematical and coding benchmarks, such as GPQA and MATH-500.
- QwQ's introspective process promotes breakthroughs in problem-solving.
- Ongoing research aims to deepen understanding of reasoning in AI.

🔖 Keywords:
#QwQ #AI #Reasoning #Learning #Mathematics

Been playing with the new QwQ 32B open weight model. Supposedly it beats OpenAI o1 and Claude Sonnet for complex reasoning for coding and maths, with a freely available model only a tiny fraction of the size.

Anyone can run it, it's good at complex code and math problems but.... I wouldn't trust a Chinese-made AI model on certain informational topics....

#qwq #artificialintelligence #AI #qwen #genai #censorship