alberthamva18/rank
1
0
Code Issues Packages Projects Wiki

Add Understanding DeepSeek R1

Albertha Vogler 2025-02-10 23:51:00 +02:00
commit e62bc0397f
1 changed files with 92 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

92
Understanding-DeepSeek-R1.md Normal file

@ -0,0 +1,92 @@
<br>DeepSeek-R1 is an [open-source language](http://vodhoz38.ru) design built on DeepSeek-V3-Base that's been making waves in the [AI](http://tvojfittrener.sk) neighborhood. Not only does it match-or even surpass-OpenAI's o1 design in numerous standards, however it also includes fully MIT-licensed weights. This marks it as the very first non-OpenAI/Google design to provide strong thinking capabilities in an open and available way.<br>
<br>What makes DeepSeek-R1 especially interesting is its openness. Unlike the less-open methods from some industry leaders, DeepSeek has published a [detailed training](http://pinkspots.de) method in their paper.
The model is also incredibly cost-effective, with input tokens costing simply $0.14-0.55 per million (vs o1's $15) and output tokens at $2.19 per million (vs o1's $60).<br>
<br>Until ~ GPT-4, the typical knowledge was that better [models required](http://git.aivfo.com36000) more data and compute. While that's still valid, designs like o1 and R1 [demonstrate](https://tdafrica.com) an option: inference-time scaling through thinking.<br>
<br>The Essentials<br>
<br>The DeepSeek-R1 paper presented numerous models, but main among them were R1 and R1-Zero. Following these are a series of distilled models that, [yewiki.org](https://www.yewiki.org/User:BasilLightfoot) while intriguing, I won't go over here.<br>
<br>DeepSeek-R1 uses two significant ideas:<br>
<br>1. A multi-stage pipeline where a little set of cold-start information kickstarts the model, followed by [large-scale RL](https://testing-sru-git.t2t-support.com).
2. Group Relative Policy Optimization (GRPO), a support knowing method that counts on comparing multiple design outputs per prompt to avoid the need for a [separate critic](https://mosrite65.com).<br>
<br>R1 and R1-Zero are both thinking models. This basically means they do Chain-of-Thought before responding to. For the R1 series of models, this takes kind as thinking within a tag, before addressing with a last summary.<br>
<br>R1-Zero vs R1<br>
<br>R1-Zero uses Reinforcement Learning (RL) straight to DeepSeek-V3-Base without any [supervised fine-tuning](http://repo.bpo.technology) (SFT). RL is utilized to optimize the model's policy to maximize benefit.
R1-Zero attains exceptional accuracy but often [produces](https://sadjiroen.de) confusing outputs, such as mixing several languages in a single reaction. R1 [repairs](http://judoclubcastenaso.it) that by incorporating minimal supervised fine-tuning and numerous RL passes, which enhances both [accuracy](https://reallygood.com) and readability.<br>
<br>It is interesting how some languages might express certain concepts much better, which leads the design to select the most expressive language for the job.<br>
<br>Training Pipeline<br>
<br>The training pipeline that [DeepSeek](http://landlady.sakura.ne.jp) released in the R1 paper is profoundly fascinating. It showcases how they [produced](http://119.3.9.593000) such [strong thinking](https://www.autourdustyle.com) designs, and what you can expect from each stage. This includes the issues that the resulting [designs](https://lasacochepourlemploi.fr) from each stage have, and how they fixed it in the next stage.<br>
<br>It's intriguing that their training pipeline varies from the typical:<br>
<br>The typical training method: Pretraining on big dataset (train to forecast next word) to get the base design → monitored fine-tuning → choice tuning by means of RLHF
R1-Zero: Pretrained → RL
R1: [morphomics.science](https://morphomics.science/wiki/User:FrancesSalerno) Pretrained → Multistage training pipeline with [multiple SFT](http://prosmotr24.ru) and RL phases<br>
<br>Cold-Start Fine-Tuning: Fine-tune DeepSeek-V3-Base on a couple of thousand Chain-of-Thought (CoT) [samples](https://theedubook.com) to guarantee the RL procedure has a decent beginning point. This provides a great design to begin RL.
First RL Stage: Apply GRPO with [rule-based rewards](https://lesdelicesdelavie.com) to improve thinking accuracy and [formatting](https://www.hungrypediaindo.com) (such as [forcing chain-of-thought](https://appsmarina.com) into [thinking](https://mysoshal.com) tags). When they were near convergence in the RL process, they moved to the next step. The outcome of this step is a strong reasoning design but with weak general abilities, e.g., poor format and language blending.
Rejection Sampling + general data: Create [brand-new SFT](https://matachot.co.il) data through [rejection tasting](http://festatable.com) on the RL checkpoint (from step 2), combined with [monitored data](https://www.travelalittlelouder.com) from the DeepSeek-V3-Base design. They gathered around 600k premium thinking samples.
Second Fine-Tuning: Fine-tune DeepSeek-V3-Base again on 800k total samples (600k thinking + 200k general tasks) for more [comprehensive capabilities](https://rathgarjuniorschool.ie). This action led to a strong thinking model with general abilities.
Second RL Stage: Add more reward signals (helpfulness, harmlessness) to fine-tune the final design, in addition to the thinking rewards. The outcome is DeepSeek-R1.
They likewise did model distillation for several Qwen and Llama designs on the reasoning traces to get distilled-R1 designs.<br>
<br>[Model distillation](http://www.papasroofing.com) is a strategy where you use an instructor model to enhance a trainee design by generating training data for the trainee design.
The instructor is usually a bigger design than the trainee.<br>
<br>Group Relative Policy Optimization (GRPO)<br>
<br>The fundamental idea behind utilizing reinforcement knowing for LLMs is to fine-tune the model's policy so that it naturally produces more accurate and beneficial responses.
They utilized a reward system that inspects not just for correctness but also for appropriate formatting and language consistency, so the model slowly finds out to favor actions that meet these quality requirements.<br>
<br>In this paper, they [encourage](https://www.pawpawzoo.com) the R1 design to produce chain-of-thought [reasoning](https://www.letsgodosomething.org) through RL training with GRPO.
Rather than including a [separate module](http://landlady.sakura.ne.jp) at inference time, the [training procedure](http://whai.space3000) itself nudges the model to produce detailed, detailed outputs-making the [chain-of-thought](https://cclofts.com) an emerging habits of the optimized policy.<br>
<br>What makes their approach especially fascinating is its reliance on straightforward, rule-based reward [functions](https://recrutd.com.au).
Instead of depending upon pricey external models or human-graded examples as in standard RLHF, the RL utilized for R1 uses basic criteria: it may give a greater benefit if the answer is proper, if it follows the anticipated/ format, and if the language of the answer matches that of the prompt.
Not relying on a reward model likewise suggests you do not have to hang out and [wiki.whenparked.com](https://wiki.whenparked.com/User:LidaChallis81) effort training it, and it doesn't take memory and [calculate](http://122.51.6.973000) away from your [main model](https://wpmu.mau.se).<br>
<br>GRPO was introduced in the DeepSeekMath paper. Here's how GRPO works:<br>
<br>1. For each input timely, the model creates various responses.
2. Each reaction gets a [scalar reward](https://www.nitangourmet.cl) based upon aspects like accuracy, formatting, and [language consistency](https://topstours.com).
3. Rewards are changed relative to the group's performance, essentially determining how much better each response is compared to the others.
4. The [design updates](https://www.tonsiteweb.be) its strategy slightly to prefer actions with greater [relative advantages](https://djceokat.com). It only makes slight adjustments-using strategies like clipping and a [KL penalty-to](http://unnewsusa.com) ensure the policy does not wander off too far from its initial behavior.<br>
<br>A cool [element](https://asenquavc.com) of GRPO is its flexibility. You can use simple rule-based benefit functions-for circumstances, granting a [bonus offer](http://www.sikimira.com) when the model correctly uses the syntax-to guide the [training](https://www.suyun.store).<br>
<br>While DeepSeek utilized GRPO, you could use alternative approaches instead (PPO or PRIME).<br>
<br>For those aiming to dive much deeper, Will Brown has actually composed rather a nice execution of training an LLM with [RL utilizing](http://120.77.205.309998) GRPO. GRPO has actually also currently been added to the Transformer Reinforcement Learning (TRL) library, [bytes-the-dust.com](https://bytes-the-dust.com/index.php/User:TamiEspinosa5) which is another good resource.
Finally, Yannic Kilcher has a terrific [video explaining](https://www.cortedeidonno.it) GRPO by going through the [DeepSeekMath paper](http://183.221.101.893000).<br>
<br>Is RL on LLMs the path to AGI?<br>
<br>As a final note on explaining DeepSeek-R1 and the approaches they have actually presented in their paper, I desire to [highlight](https://djceokat.com) a passage from the DeepSeekMath paper, based upon a point Yannic Kilcher made in his video.<br>
<br>These findings suggest that RL boosts the model's overall performance by rendering the output distribution more robust, in other words, it appears that the is attributed to improving the correct response from TopK rather than the enhancement of fundamental capabilities.<br>
<br>In other words, RL fine-tuning tends to form the output circulation so that the highest-probability outputs are more likely to be right, despite the fact that the total capability (as determined by the diversity of appropriate answers) is mainly present in the pretrained design.<br>
<br>This suggests that [support learning](http://zxos.vip) on LLMs is more about refining and "shaping" the existing circulation of reactions instead of endowing the model with entirely new abilities.
Consequently, while RL methods such as PPO and GRPO can produce considerable performance gains, there seems an inherent ceiling figured out by the underlying model's pretrained knowledge.<br>
<br>It is uncertain to me how far RL will take us. Perhaps it will be the stepping stone to the next huge milestone. I'm excited to see how it unfolds!<br>
<br>Running DeepSeek-R1<br>
<br>I've used DeepSeek-R1 through the main chat user interface for different issues, which it seems to solve all right. The extra search functionality makes it even nicer to utilize.<br>
<br>Interestingly, o3-mini(-high) was launched as I was composing this post. From my preliminary testing, R1 appears more powerful at [mathematics](http://thehotelandrea.com) than o3-mini.<br>
<br>I likewise rented a single H100 through Lambda Labs for $2/h (26 CPU cores, 214.7 GB RAM, 1.1 TB SSD) to run some experiments.
The main goal was to see how the design would [perform](http://srtroyfact.ru) when [deployed](https://video.salamalikum.com) on a single H100 GPU-not to thoroughly test the model's abilities.<br>
<br>671B through Llama.cpp<br>
<br>DeepSeek-R1 1.58-bit (UD-IQ1_S) quantized model by Unsloth, with a 4-bit quantized KV-cache and partial GPU [offloading](https://www.sebastiapons.com) (29 layers working on the GPU), [historydb.date](https://historydb.date/wiki/User:Tommie1225) running through llama.cpp:<br>
<br>29 layers appeared to be the sweet spot given this [configuration](https://gitlab.iue.fh-kiel.de).<br>
<br>Performance:<br>
<br>A r/[localllama](https://tramadol-online.org) user explained that they had the ability to get over 2 tok/sec with DeepSeek R1 671B, without [utilizing](https://evoluaclinica.com.br) their GPU on their local video gaming setup.
Digital Spaceport composed a complete guide on how to run Deepseek R1 671b fully locally on a $2000 EPYC server, on which you can get ~ 4.25 to 3.5 tokens per second. <br>
<br>As you can see, the tokens/s isn't rather manageable for any major work, however it's fun to run these large models on available hardware.<br>
<br>What matters most to me is a combination of effectiveness and time-to-usefulness in these models. Since reasoning models need to believe before responding to, their time-to-usefulness is generally higher than other designs, however their usefulness is also normally higher.
We require to both take full advantage of effectiveness and lessen time-to-usefulness.<br>
<br>70B by means of Ollama<br>
<br>70.6 b params, 4-bit KM [quantized](http://live.lodman.pl) DeepSeek-R1 running by means of Ollama:<br>
<br>[GPU utilization](https://www.loftcommunications.com) soars here, as anticipated when compared to the mainly CPU-powered run of 671B that I showcased above.<br>
<br>Resources<br>
<br>DeepSeek-R1: Incentivizing Reasoning Capability in LLMs through [Reinforcement Learning](http://argonizer.ru)
[2402.03300] DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
DeepSeek R1 - Notion (Building a totally local "deep researcher" with DeepSeek-R1 - YouTube).
DeepSeek R1's dish to replicate o1 and the future of reasoning LMs.
The Illustrated DeepSeek-R1 - by [Jay Alammar](https://www.moodswingsmusic.nl).
Explainer: What's R1 & Everything Else? - Tim [Kellogg](http://umfp.ma).
DeepSeek R1 Explained to your granny - YouTube<br>
<br>DeepSeek<br>
<br>- Try R1 at [chat.deepseek](http://digmbio.com).com.
GitHub - deepseek-[ai](http://baarn.co.kr)/DeepSeek-R 1.
deepseek-[ai](https://gneistspelen.gneist.org)/Janus-Pro -7 B [· Hugging](https://zakirov-prod.ru) Face (January 2025): [morphomics.science](https://morphomics.science/wiki/User:BerryY9555701703) Janus-Pro is a novel autoregressive structure that combines multimodal understanding and generation. It can both comprehend and produce images.
DeepSeek-R1: Incentivizing Reasoning Capability in Large [Language](https://shoden-giken.com) Models by means of Reinforcement Learning (January 2025) This paper presents DeepSeek-R1, an open-source reasoning design that rivals the efficiency of [OpenAI's](https://95theses.co.uk) o1. It provides a detailed methodology for training such designs using [massive support](https://www.surkhab7.com) knowing strategies.
DeepSeek-V3 Technical Report (December 2024) This report discusses the implementation of an FP8 blended accuracy [training framework](http://www.marvelcompany.co.jp) confirmed on an extremely large-scale design, attaining both sped up training and [lowered GPU](https://transparencia.ahome.gob.mx) memory usage.
DeepSeek LLM: Scaling Open-Source Language Models with Longtermism (January 2024) This paper dives into scaling laws and provides findings that assist in the scaling of massive models in open-source configurations. It presents the DeepSeek LLM project, dedicated to advancing open-source language designs with a long-lasting perspective.
DeepSeek-Coder: When the Large Language Model Meets Programming-The Rise of [Code Intelligence](https://vailmillrace.com) (January 2024) This research study introduces the [DeepSeek-Coder](https://idawulff.no) series, a range of open-source code models trained from scratch on 2 trillion tokens. The designs are pre-trained on a top quality project-level code corpus and use a fill-in-the-blank job to improve code generation and infilling.
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model (May 2024) This paper provides DeepSeek-V2, a Mixture-of-Experts (MoE) language model characterized by cost-effective training and efficient inference.
DeepSeek-Coder-V2: Breaking the Barrier of [Closed-Source Models](http://allncorp.com) in Code Intelligence (June 2024) This research introduces DeepSeek-Coder-V2, an [open-source Mixture-of-Experts](https://www.michellemareesschoolofdance.com.au) (MoE) [code language](https://projobs.dk) design that attains performance [equivalent](http://srtroyfact.ru) to GPT-4 Turbo in code-specific jobs.<br>
<br>Interesting events<br>
<br>- Hong Kong [University duplicates](https://adel-watch.de) R1 outcomes (Jan 25, '25).
- Huggingface [reveals](https://englishfunclub.pl) huggingface/open-r 1: Fully open reproduction of DeepSeek-R1 to reproduce R1, totally open source (Jan 25, [forum.pinoo.com.tr](http://forum.pinoo.com.tr/profile.php?id=1323760) '25).
- OpenAI researcher confirms the DeepSeek group independently found and [utilized](https://www.rallydecoracoes.com.br) some [core ideas](https://rainer-transport.com) the OpenAI group utilized en route to o1<br>
<br>Liked this post? Join the newsletter.<br>