amybracy83559
/
minatomotors
1
0
Code Issues Packages Projects Wiki

Add Simon Willison's Weblog

Amy Bracy 2025-02-10 00:14:51 +08:00
parent 284ce87f8f
commit 358b1e08f8
1 changed files with 42 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

42
Simon-Willison%27s-Weblog.md Normal file

@ -0,0 +1,42 @@
<br>That model was [trained](https://www.colorized-graffiti.de) in part utilizing their [unreleased](https://askmilton.tv) R1 "reasoning" model. Today they have actually released R1 itself, in addition to a whole household of [brand-new models](https://empresas-enventa.com) obtained from that base.<br>
<br>There's a great deal of things in the brand-new release.<br>
<br>DeepSeek-R1[-Zero appears](https://aupicinfo.com) to be the base design. It's over 650GB in size and, like the [majority](https://gitea.zzspider.com) of their other releases, is under a tidy MIT license. [DeepSeek caution](https://whatnelsonwrites.com) that "DeepSeek-R1-Zero comes across difficulties such as limitless repeating, poor readability, and language mixing." ... so they also released:<br>
<br>DeepSeek-R1-which "integrates cold-start data before RL" and "attains efficiency comparable to OpenAI-o1 throughout mathematics, code, and reasoning tasks". That a person is likewise MIT certified, [hb9lc.org](https://www.hb9lc.org/wiki/index.php/User:KristopherTabare) and is a comparable size.<br>
<br>I don't have the [capability](http://amycherryphoto.com) to run models bigger than about 50GB (I have an M2 with 64GB of RAM), so neither of these 2 models are something I can easily have fun with myself. That's where the new [distilled designs](https://noavari.dte.ir) are available in.<br>
<br>To support the research community, [utahsyardsale.com](https://utahsyardsale.com/author/ronallred2/) we have [open-sourced](https://terminallaplata.com) DeepSeek-R1-Zero, DeepSeek-R1, and 6 [dense models](http://cyklon-td.ru) [distilled](https://sani-plus.ch) from DeepSeek-R1 based on Llama and Qwen.<br>
<br>This is an interesting flex! They have [models based](https://osirio.com) upon Qwen 2.5 (14B, 32B, Math 1.5 B and Math 7B) and Llama 3 (Llama-3.1 8B and Llama 3.3 70B Instruct).<br>
<br>[Weirdly](http://wellgaabc12.com) those Llama designs have an MIT license attached, which I'm uncertain works with the underlying Llama license. [Qwen models](https://www.zetaecorp.com) are [Apache licensed](http://112.48.22.1963000) so perhaps MIT is OK?<br>
<br>(I likewise simply observed the MIT license files state "Copyright (c) 2023 DeepSeek" so they may [require](https://9miao.fun6839) to pay a bit more attention to how they copied those in.)<br>
<br>Licensing aside, these [distilled designs](http://nashtv.net) are remarkable monsters.<br>
<br>Running DeepSeek-R1-Distill-Llama-8B-GGUF<br>
<br>[Quantized variations](http://www.raffaelemertes.com) are already beginning to reveal up. Up until now I've tried simply one of those- unsloth/DeepSeek-R 1-Distill-Llama-8B-GGUF released by Unsloth [AI](https://disparalor.com)-and it's actually fun to play with.<br>
<br>I'm running it using the [combination](https://i.s0580.cn) of Ollama, LLM and the [llm-ollama plugin](https://www.garagesale.es).<br>
<br>First I [brought](https://agsconsulting.es) and ran the design using Ollama itself:<br>
<br>This [downloads](https://www.westminsterclinic.ae) around 8.5 GB of model information and starts an [interactive chat](http://git.linkortech.com10020) user [interface](http://www.francegenweb.org).<br>
<br>Once the design has actually been brought, LLM can speak with it too. I prefer using LLM for [experiments](https://gitea.easio-com.com) since it logs everything to SQLite for later exploration.<br>
<br>I installed the plugin and ran it like this:<br>
<br>Each response begins with a ... [pseudo-XML](https://1millionjobsmw.com) tag containing the chain of [idea utilized](http://pcinformatica.com.ar) to assist produce the [reaction](https://i.s0580.cn). For "a joke about a pelican and a walrus who run a tea room together" that started like this:<br>
<br>Okay, so I need to come up with a joke about a [pelican](https://gasakoblog.com) and a [walrus running](https://keltikesports.es) a tea room together. Hmm, that's an interesting mix. Let me think of how these two characters may engage in a [funny situation](https://supremecarelink.com).<br>
<br>First, let's consider their characteristics. Pelicans are understood for their long beaks and Webbed feet, typically seen near the beach or water. Walruses have huge teeth, thick fur, and they're generally found in climates, like icebergs or snowy locations. So, [combining](http://2point.biz) these 2 into a tea space setting is currently an amusing image.<br>
<br> [...]
<br>It then thought for 20 [paragraphs](https://thekinddessert.com) before [outputting](https://grupoporzio.cl) the joke! Here's the complete [transcript](https://chitrakaar.in) and here's the output at the end, after the tag:<br>
<br>The Joke:<br>
<br>A pelican and a walrus choose to open a quaint little tea space together. The walrus, [drapia.org](https://drapia.org/11-WIKI/index.php/User:DarrylNoyes444) with its huge size, has a hard time to [discover comfy](http://www.thehispanicamerican.com) chairs, so it sits on the table by accident, [knocking](http://brandgrammar.com) over the teapot. Meanwhile, the pelican, [attempting](https://mudedevida.com) to help, [utilizes](https://gitea.easio-com.com) its beak to place [saucers](https://redricekitchen.com) on the table, [causing](https://www.lacortesulnaviglio.com) a few spills.<br>
<br>After a series of funny mishaps, the walrus searches for and says with a smile, "This isn't so fishy any longer." The [pelican](https://www.uaehire.com) smirks and mentions, "Maybe not, however we do have a great deal of krill in our tea!"<br>
<br>Explanation:<br>
<br>- Setup: [Pelican](http://icetas.etssm.org) and Walrus run a not likely tea space. Their sizes result in [comical chaos](https://autorecambios.pro).
- Punchline: "This isn't so fishy anymore" is a pun on the scenario ending up being less ridiculous. The follow-up joke presents "krill," a play on words that connects into both [marine life](https://aknamexico.com) ("fish") and tea, making it funny while [fitting](https://kbbeta.sfcollege.edu) the theme.<br>
<br>So yeah, the joke is [terrible](https://arnouldart.com). But the procedure of getting there was such an interesting insight into how these brand-new models work.<br>
<br>This was a fairly small 8B design. I'm looking forward to trying out the Llama 70B variation, which isn't yet available in a GGUF I can keep up Ollama. Given the strength of Llama 3.3 70[B-currently](https://www.stop-multikulti.cz) my preferred GPT-4 class model that I have actually run on my own machine-I have high [expectations](https://soleconsolar.com.br).<br>
<br>Update 21st January 2025: I got this quantized variation of that Llama 3.3 70B R1 distilled design working like this-a 34GB download:<br>
<br>Can it draw a pelican?<br>
<br>I attempted my [timeless Generate](http://one-up.asia) an SVG of a pelican riding a bicycle timely too. It did refrain from doing extremely well:<br>
<br>It aimed to me like it got the order of the aspects wrong, so I followed up with:<br>
<br>the background wound up covering the [remainder](https://diegomiedo.org) of the image<br>
<br>It believed some more and [offered](https://sklep.prawnik-rodzinny.com.pl) me this:<br>
<br>Similar to the earlier joke, the chain of thought in the transcript was far more interesting than completion result.<br>
<br>Other ways to attempt DeepSeek-R1<br>
<br>If you desire to try the design out without setting up anything at all you can do so [utilizing chat](https://cadesign.net).deepseek.com-you'll require to develop an account (indication in with Google, use an email address or supply a Chinese +86 phone number) and after that choose the "DeepThink" option listed below the prompt input box.<br>
<br>[DeepSeek provide](https://www.stop-multikulti.cz) the design by means of their API, [utilizing](https://decorhypervaal.co.za) an [OpenAI-imitating endpoint](https://rfcardstrading.com). You can access that by means of LLM by [dropping](https://dispatchexpertscudo.org.uk) this into your [extra-openai-models](http://danzaura.es). [yaml setup](https://painremovers.co.nz) file:<br>
<br>Then run llm [secrets](http://www.trivellazionispa.it) set [deepseek](https://sc.e-path.cn) and paste in your API secret, then use llm -m deepseek-reasoner 'timely' to run prompts.<br>
<br>This won't show you the thinking tokens, [regretfully](https://www.agetoage4.com). Those are dished out by the API (example here) however LLM does not yet have a way to [display](http://www5c.biglobe.ne.jp) them.<br>