diff --git a/DeepSeek-R1%3A-Technical-Overview-of-its-Architecture-And-Innovations.md b/DeepSeek-R1%3A-Technical-Overview-of-its-Architecture-And-Innovations.md index 17ea9c6..ee9de2f 100644 --- a/DeepSeek-R1%3A-Technical-Overview-of-its-Architecture-And-Innovations.md +++ b/DeepSeek-R1%3A-Technical-Overview-of-its-Architecture-And-Innovations.md @@ -1,54 +1,54 @@ -
DeepSeek-R1 the latest [AI](https://tomeknawrocki.pl) design from Chinese start-up [DeepSeek](http://www.ads-chauffeur.fr) [represents](http://florence.boignard.free.fr) a cutting-edge advancement in [generative](http://au-elista.ru) [AI](https://www.fotopaletti.it) innovation. [Released](https://grupormk.com) in January 2025, it has [gained worldwide](https://gitea.echocolate.xyz) [attention](https://www.jerseylawoffice.com) for its innovative architecture, cost-effectiveness, and extraordinary performance across [multiple domains](https://winatlifeli.org).
+
DeepSeek-R1 the most recent [AI](http://www.gnovi.in) design from [Chinese startup](https://masonhardwareuk.co.uk) [DeepSeek represents](https://sportysocialspace.com) a [groundbreaking development](https://mponlinecoaching.pt) in generative [AI](https://gitea.viewdeco.cn) technology. [Released](https://468innovation.com) in January 2025, it has [gained global](https://guitaration.com) attention for its [ingenious](https://yu2ube.com) architecture, cost-effectiveness, and [remarkable efficiency](http://www.bsr-secure.eu) throughout several [domains](https://www.comforttime.net).

What Makes DeepSeek-R1 Unique?
-
The [increasing demand](https://www.space2b.org.uk) for [AI](http://sinbiromall.hubweb.net) [models efficient](https://sanantoniohailclaims.com) in dealing with intricate reasoning jobs, long-context understanding, and domain-specific versatility has [exposed](https://maverick-services.com.sg) constraints in traditional dense transformer-based designs. These models frequently experience:
-
High computational costs due to activating all parameters throughout inference. -
[Inefficiencies](https://www.gameenthus.com) in multi-domain job handling. -
Limited scalability for large-scale implementations. -
-At its core, DeepSeek-R1 differentiates itself through a powerful mix of scalability, efficiency, and high efficiency. Its architecture is developed on two [foundational](http://turbocharger.ru) pillars: a cutting-edge Mixture of Experts (MoE) [structure](https://www.x-shai.com) and an [advanced transformer-based](http://lvps83-169-32-176.dedicated.hosteurope.de) design. This hybrid technique enables the model to take on complicated tasks with exceptional accuracy and speed while [maintaining cost-effectiveness](https://kwenenggroup.com) and [attaining cutting](https://instashare.net) edge outcomes.
-
Core Architecture of DeepSeek-R1
-
1. [Multi-Head Latent](https://kpi-eg.ru) [Attention](http://git.cyjyyjy.com) (MLA)
-
MLA is a [crucial architectural](https://vlauncher.net) innovation in DeepSeek-R1, presented at first in DeepSeek-V2 and further [refined](http://williammcgowanlettings.com) in R1 developed to [enhance](https://www.habreha.nl) the [attention](http://julietteduprez-psychotherapie.fr) system, minimizing memory overhead and computational inefficiencies throughout reasoning. It runs as part of the [design's core](https://www.premium-english.pl) architecture, straight affecting how the model procedures and creates outputs.
-
[Traditional multi-head](https://liquidmixagitators.com) attention calculates separate Key (K), Query (Q), and Value (V) matrices for each head, which scales quadratically with input size. -
[MLA replaces](http://www.ethansoloviev.com) this with a [low-rank factorization](http://territoriyapodarkov.ru) method. Instead of [caching](http://dagatron.com) complete K and V matrices for each head, [MLA compresses](https://www.colegiocaminoabelen.com) them into a hidden vector. -
-During inference, these hidden vectors are decompressed on-the-fly to recreate K and V [matrices](https://www.afrigodigit.com) for each head which [considerably lowered](https://cleanbyjolene.com) KV-cache size to just 5-13% of [conventional techniques](http://kosmosgida.com).
-
Additionally, MLA [integrated Rotary](http://www.xn--80aafblbgpxxcgbigyfoeei.xn--p1ai) [Position Embeddings](http://121.181.234.77) (RoPE) into its design by [devoting](https://www.crivian2.it) a part of each Q and K head specifically for positional details avoiding [redundant knowing](http://demo.amytheme.com) across heads while maintaining compatibility with [position-aware tasks](https://greenhedgehog.at) like long-context thinking.
-
2. [Mixture](https://git.adminkin.pro) of Experts (MoE): The [Backbone](http://www.daytonaraceurope.eu) of Efficiency
-
MoE structure permits the design to dynamically activate just the most appropriate sub-networks (or "professionals") for a given job, guaranteeing effective resource utilization. The architecture consists of 671 billion parameters distributed throughout these [professional networks](https://rootwholebody.com).
-
Integrated dynamic gating system that does something about it on which experts are activated based upon the input. For any given query, just 37 billion criteria are activated during a single forward pass, significantly decreasing computational overhead while [maintaining](http://www.ads-chauffeur.fr) high performance. -
This [sparsity](https://mptradio.com) is attained through methods like Load Balancing Loss, which guarantees that all [specialists](https://awaregift.com) are made use of uniformly over time to avoid traffic jams. -
-This architecture is built upon the structure of DeepSeek-V3 (a pre-trained foundation model with robust general-purpose capabilities) even more refined to [boost thinking](https://primetimecommentary.com) abilities and [domain adaptability](https://the24watch.shop).
-
3. Transformer-Based Design
-
In addition to MoE, DeepSeek-R1 [incorporates advanced](http://maitri.adaptiveit.net) transformer layers for natural language processing. These layers incorporates optimizations like sporadic attention systems and [effective tokenization](https://comercialmym.cl) to catch contextual relationships in text, enabling remarkable [comprehension](http://wadfotografie.nl) and action generation.
-
[Combining hybrid](https://apkjobs.com) attention mechanism to dynamically [adjusts attention](https://gingerbread-mansion.com) weight distributions to [optimize](http://evasampe-cp43.wordpresstemporal.com) [efficiency](https://ibizabouff.be) for both short-context and long-context circumstances.
-
[Global Attention](https://www.gameenthus.com) records [relationships](http://tb1561.nyuad.im) across the entire input sequence, perfect for jobs requiring [long-context](https://www.ttg.cz) understanding. -
Local Attention [focuses](https://git.ivran.ru) on smaller, contextually considerable segments, such as adjacent words in a sentence, [improving effectiveness](http://matzkemedia.de) for language tasks. -
-To simplify input processing [advanced tokenized](https://www.afrigodigit.com) methods are incorporated:
-
Soft Token Merging: [merges redundant](http://karboglass18.ru) tokens during processing while [maintaining critical](https://vlauncher.net) details. This decreases the number of tokens gone through [transformer](http://www.thenghai.org.sg) layers, [improving computational](https://casopis.feb.ba) performance -
Dynamic Token Inflation: counter prospective [details loss](https://tkmwp.com) from token combining, the [design utilizes](https://khatmedun.tj) a token inflation module that restores essential details at later [processing](http://devcons.ro) stages. -
-[Multi-Head Latent](http://geek-leak.com) [Attention](https://niktalkmedia.com) and [Advanced](http://giwa.shop) Transformer-Based Design are [carefully](https://git.jaronnie.com) associated, as both handle attention systems and transformer architecture. However, they concentrate on various aspects of the architecture.
-
MLA particularly [targets](http://caspian-baku-logistic.com) the [computational performance](https://aroapress.com) of the attention mechanism by [compressing](http://www.grupsa.in) Key-Query-Value (KQV) matrices into latent areas, decreasing memory overhead and [inference latency](https://git.tintinger.org). -
and Advanced Transformer-Based Design concentrates on the general optimization of transformer layers. -
-[Training Methodology](https://tubeseen.com) of DeepSeek-R1 Model
-
1. [Initial Fine-Tuning](http://schietverenigingterschuur.nl) (Cold Start Phase)
-
The begins with [fine-tuning](http://121.181.234.77) the base design (DeepSeek-V3) using a little dataset of carefully curated chain-of-thought (CoT) thinking examples. These [examples](https://www.otusagenciadigital.com.br) are thoroughly curated to guarantee variety, clarity, and sensible [consistency](http://124.70.145.1510880).
-
By the end of this phase, the model demonstrates improved thinking capabilities, setting the stage for more innovative training stages.
+
The increasing need for [AI](http://modestecorrecteur.blog.free.fr) models capable of dealing with intricate reasoning jobs, long-context understanding, and domain-specific versatility has actually [exposed](https://www.top5stockbroker.com) constraints in conventional dense transformer-based models. These models often with:
+
High computational costs due to [triggering](http://www.pehlivanogluyapi.com) all parameters during [reasoning](http://lanciaaustralia.com.au). +
Inefficiencies in [multi-domain job](https://www.bedasso.org.uk) handling. +
Limited scalability for massive releases. +
+At its core, DeepSeek-R1 distinguishes itself through an [effective combination](http://47-1.eu) of scalability, effectiveness, and high [efficiency](http://valueadd.kr). Its [architecture](http://interdecorpro.pl) is constructed on 2 [fundamental](https://matachot.co.il) pillars: an advanced Mixture of [Experts](https://batfriendly.org) (MoE) [framework](http://wrs.spdns.eu) and an [innovative transformer-based](https://git.declic3000.com) style. This [hybrid approach](https://gitlab.cheretech.com) [permits](https://www.covoiturage.cm) the design to [tackle complicated](https://hoathinhvn.com) jobs with [remarkable accuracy](https://livingspaces.ie) and speed while maintaining cost-effectiveness and [attaining](https://www.sustainablewaterlooregion.ca) modern results.
+
[Core Architecture](http://gitlab.hiperpbx.com) of DeepSeek-R1
+
1. [Multi-Head](https://kyno.network) [Latent Attention](https://virtualdata.pt) (MLA)
+
MLA is a [vital architectural](http://pragmatikcozumler.com) innovation in DeepSeek-R1, presented [initially](https://pelangideco.com) in DeepSeek-V2 and more improved in R1 created to [optimize](http://www.seamlessnc.com) the attention system, minimizing memory [overhead](http://inmoportal.es) and [computational](https://git.hackercan.dev) inadequacies throughout inference. It [operates](https://mojecoventry.pl) as part of the [design's core](https://puertanatura.es) architecture, [straight impacting](https://denkleinedommel.nl) how the design processes and creates outputs.
+
[Traditional multi-head](https://purcolor.at) [attention computes](https://congxepgiatung.com) different Key (K), Query (Q), and Value (V) [matrices](https://gravesmediagroup.com) for each head, which scales quadratically with [input size](https://dobetterhub.com). +
[MLA replaces](https://tridentbuildingandroofing.co.uk) this with a [low-rank](https://www.eshoplogistic.com) factorization technique. Instead of [caching](http://www.erlingtingkaer.dk) full K and V [matrices](https://bouticar.com) for each head, [MLA compresses](https://www.homecookingwithkimberly.com) them into a [latent vector](https://www.ahhand.com). +
+During reasoning, these latent vectors are [decompressed](https://free-weblink.com) [on-the-fly](https://shandeeland.com) to [recreate K](https://mozillabd.science) and V [matrices](https://jph.dk) for each head which significantly [lowered KV-cache](http://digimc.co) size to just 5-13% of [conventional](https://rlt.com.np) approaches.
+
Additionally, [MLA integrated](http://gac-cont.com) [Rotary Position](https://vloglover.com) [Embeddings](https://bahnreise-wiki.de) (RoPE) into its style by [dedicating](http://khk.co.ir) a part of each Q and K head particularly for positional details [avoiding](https://www.oemautomation.com8888) [redundant learning](https://www.noagagu.kr) across heads while [maintaining compatibility](https://www.ahhand.com) with position-aware tasks like [long-context](https://www.thesevenoaksanimator.com) reasoning.
+
2. Mixture of [Experts](https://mixto.ro) (MoE): The Backbone of Efficiency
+
MoE framework [enables](https://christianbiz.ca) the design to [dynamically trigger](http://www.artsphera.com.ua) just the most [pertinent sub-networks](https://sandrapronkinterim.nl) (or "professionals") for a provided job, [guaranteeing](https://gruposanvicentegalapagos.com) [efficient resource](https://pogruz.kg) usage. The architecture consists of 671 billion [parameters distributed](http://thesplendidlifestyle.com) across these professional networks.
+
Integrated [vibrant gating](http://digitalmarketingconnection.com) system that acts on which [professionals](https://grand.parts) are [triggered based](https://hospitalitymatches.com) on the input. For any given inquiry, only 37 billion [criteria](http://yamarashi.it) are [activated](https://brownscakes.com) during a single forward pass, [considerably lowering](https://406.gotele.net) computational overhead while [maintaining](https://www.kaokimhourn.com) high performance. +
This [sparsity](https://virtualdata.pt) is attained through strategies like Load Balancing Loss, which ensures that all [specialists](https://www.virsistance.com) are made use of evenly [gradually](http://369ant.com) to [prevent traffic](http://carecall.co.kr) jams. +
+This [architecture](http://sq.alssunnah.com) is [developed](https://www.haskinlawoakcreek.com) upon the structure of DeepSeek-V3 (a [pre-trained structure](https://erwincaubergh.be) design with [robust general-purpose](https://jonaogroup.com) capabilities) even more [refined](https://.ob.ejam.esa.le.ngjianf.ei2013%25252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252528...252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252529a.langtonSus.ta.i.n.j.ex.kfen.gku.an.gx.r.ku.ai8.xn252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252520.xn252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252520.u.kMeli.s.a.ri.c.h4223e.xultan.tacoustic.sfat.lettuceerzfault.ybeamdulltnderwearertwe.s.ep.laus.i.bleljhr.eces.si.v.e.x.g.zleanna.langtonWww.emekaolisawww.karunakumari46sh.jdus.h.a.i.j.5.8.7.4.8574.85c.o.nne.c.t.tn.tuGo.o.gle.email.2.25252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525255c25252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525252525255cn1sarahjohnsonw.estbrookbertrew.e.rhu.fe.ng.k.ua.ngniu.bi..uk41Www.zanelesilvia.woodw.o.r.t.hw.anting.parentcrazyre.stfir.stdrowww.mondaymorninginspirationfidelia.commonsHu.fen.gk.uang.ni.u.b.i.xn--.u.k.6.2p.a.r.a.ju.mp.e.r.sj.a.s.s.en20.14Leanna.langtonYour.qwe.aqmailSus.ta.i.n.j.ex.kwww.darccycling.com) to [improve reasoning](https://greenpeacefoundation.com) abilities and [domain flexibility](https://blackroommedia.com).
+
3. [Transformer-Based](https://sharefolks.com) Design
+
In addition to MoE, DeepSeek-R1 integrates innovative [transformer](https://kyno.network) layers for [natural](https://lambdahub.yavin4.ch) [language](https://bulletproof-media.com) processing. These layers incorporates [optimizations](https://gitea.viewdeco.cn) like sporadic attention systems and efficient tokenization to catch contextual relationships in text, making it possible for exceptional comprehension and [response generation](https://www.bitontocortiliaperti.it).
+
Combining hybrid attention mechanism to [dynamically adjusts](https://gitea.viewdeco.cn) attention weight distributions to optimize performance for both short-context and long-context scenarios.
+
Global Attention [catches](https://projobs.dk) [relationships](https://grade1d.smaportal.ae) across the entire input series, [suitable](https://durbanpainter.co.za) for [jobs requiring](http://vis.edu.in) long-context comprehension. +
Local Attention [concentrates](https://barleysmenu.com) on smaller, [contextually considerable](https://belclarefarm.com) segments, such as nearby words in a sentence, [enhancing effectiveness](http://star-ship-jpn.com) for language jobs. +
+To improve [input processing](http://www.naijanetwork.com) advanced [tokenized](https://joventic.uoc.edu) [methods](https://chalet-binii.ch) are incorporated:
+
[Soft Token](https://asiatex.fr) Merging: merges redundant tokens throughout [processing](https://beach69-kamomi.com) while maintaining vital [details](https://kzashop.com). This decreases the number of tokens passed through [transformer](http://jukatrashy.com) layers, improving computational [performance](http://d4bh.ru) +
[Dynamic Token](http://consulam.com) Inflation: [counter](https://git.guildofwriters.org) [potential details](https://jiangjianhua2525.com) loss from token combining, the design utilizes a token inflation module that restores key [details](https://myhcpna.org) at later processing phases. +
+[Multi-Head](http://www.prismaimoveisilha.com.br) [Latent Attention](http://sci-admin.org) and Advanced Transformer-Based Design are [carefully](https://michiganpipelining.com) associated, as both deal with attention mechanisms and transformer architecture. However, [lovewiki.faith](https://lovewiki.faith/wiki/User:EveretteLemmone) they focus on various aspects of the architecture.
+
MLA particularly targets the [computational effectiveness](https://www.nikisalons.com) of the [attention mechanism](http://dashausammeer.com) by [compressing](http://47.93.56.668080) [Key-Query-Value](https://www.handinhandspace.com) (KQV) matrices into latent areas, [decreasing memory](https://www.haskinlawoakcreek.com) overhead and inference latency. +
and Advanced Transformer-Based Design focuses on the general [optimization](http://explodingfreedomcentralcity.shoutwiki.com) of transformer layers. +
+Training Methodology of DeepSeek-R1 Model
+
1. [Initial Fine-Tuning](https://wrapupped.com) (Cold Start Phase)
+
The procedure starts with fine-tuning the [base design](https://www.creamteasandchampagne.com) (DeepSeek-V3) utilizing a small [dataset](http://teubes.com) of carefully curated [chain-of-thought](https://kyno.network) (CoT) [reasoning examples](http://secondlinejazzband.com). These [examples](https://sk303.com) are [carefully curated](https://aubookcafe.com) to make sure diversity, clearness, and [rational consistency](https://www.wallpostjournal.com).
+
By the end of this stage, the model [demonstrates enhanced](http://s522908547.online.de) thinking capabilities, [setting](https://barrierskate.com) the phase for [advanced training](https://seatcovers.co.za) phases.

2. Reinforcement Learning (RL) Phases
-
After the [preliminary](https://www.jefffoster.net) fine-tuning, DeepSeek-R1 [undergoes multiple](http://www.ecodacs2.nerima.tokyo.jp) [Reinforcement Learning](http://opensees.ir) (RL) stages to further refine its reasoning abilities and [guarantee](http://www.rakutaku.com) [alignment](http://kingzcorner.de) with human choices.
-
Stage 1: Reward Optimization: Outputs are incentivized based on accuracy, readability, and formatting by a [reward design](https://tejgujarati.com). -
Stage 2: Self-Evolution: Enable the design to [autonomously develop](https://electrilight.ca) [advanced thinking](https://bannermation.co.uk) habits like [self-verification](https://magikos.sk) (where it examines its own [outputs](https://gitea.mpc-web.jp) for [consistency](https://openerp.vn) and correctness), [reflection](https://skylift.gr) (recognizing and correcting mistakes in its reasoning procedure) and mistake correction (to [improve](http://spnewstv.com) its [outputs iteratively](http://www.daytonaraceurope.eu) ). -
Stage 3: Helpfulness and [Harmlessness](https://workbook.ai) Alignment: Ensure the [design's outputs](http://119.45.49.2123000) are useful, safe, and lined up with human choices. +
After the [preliminary](https://www.birreriareartu.com) fine-tuning, DeepSeek-R1 goes through [numerous Reinforcement](https://demo.pixelphotoscript.com) [Learning](https://www.pathwayfc.org) (RL) stages to further [fine-tune](https://placementug.com) its [reasoning capabilities](https://markfedpunjab.com) and make sure [positioning](https://bundas24.com) with [human choices](https://a2b.ba).
+
Stage 1: Reward Optimization: [Outputs](https://linkspreed.web4.one) are incentivized based on precision, readability, and [formatting](https://medicinudenrecept.com) by a [reward design](https://mantovauno.it). +
Stage 2: Self-Evolution: Enable the model to autonomously develop [sophisticated](https://erwincaubergh.be) [reasoning behaviors](https://somayehtrading.com) like self-verification (where it [inspects](https://www.covoiturage.cm) its own [outputs](https://personal.spaces.one) for [consistency](https://ermatorusa.com) and correctness), [reflection](https://barleysmenu.com) ([recognizing](https://hostalcalaratjada.com) and [fixing errors](https://chalet-binii.ch) in its [thinking](https://www.bailados.com.br) procedure) and [error correction](https://quickservicesrecruits.com) (to refine its outputs iteratively ). +
Stage 3: [Helpfulness](https://aubookcafe.com) and [Harmlessness](https://climbelectric.com) Alignment: Ensure the [design's outputs](https://bestadjustablebeds.net) are useful, safe, and aligned with human preferences.
-3. Rejection [Sampling](https://www.comesuomo1974.com) and Supervised Fine-Tuning (SFT)
-
After creating a great deal of samples just top quality outputs those that are both accurate and readable are selected through [rejection sampling](http://havefotografi.dk) and reward model. The model is then further trained on this fine-tuned dataset utilizing [supervised](https://srapo.com) fine-tuning, which [consists](http://nongtachiang.ssk.in.th) of a wider series of questions beyond [reasoning-based](https://www.promotstore.com) ones, boosting its [efficiency](http://anytimefitness-ek.co.uk) across [numerous domains](https://kaiftravels.com).
+3. [Rejection Sampling](https://paradig.eu) and [Supervised Fine-Tuning](https://mojecoventry.pl) (SFT)
+
After [producing](http://www.avtoshkola63.ru) a great deal of samples just [premium outputs](https://epe31.fr) those that are both precise and readable are [selected](https://hookahtobaccogermany.de) through rejection sampling and [reward design](http://mijutech.com). The design is then [additional trained](http://synaps-audiovisuel.fr) on this improved dataset [utilizing](http://consulam.com) monitored fine-tuning, which includes a more [comprehensive series](https://terryhobbs.com) of concerns beyond reasoning-based ones, [enhancing](https://professionpartners.co.uk) its [proficiency](https://refidomsa.hubmoe.com) across several [domains](http://82.19.55.40443).

Cost-Efficiency: A Game-Changer
-
DeepSeek-R1['s training](https://lungnancy11.edublogs.org) cost was around $5.6 million-significantly lower than [competing designs](https://koureisya.com) trained on costly Nvidia H100 GPUs. [Key factors](https://en-musubi-yukari.com) contributing to its [cost-efficiency consist](https://clujjobs.com) of:
-
MoE architecture decreasing computational requirements. -
Use of 2,000 H800 GPUs for [drapia.org](https://drapia.org/11-WIKI/index.php/User:IssacHicks62581) training instead of higher-cost alternatives. +
DeepSeek-R1['s training](https://afromonsta.com) [expense](https://blueskygroup.com.au) was around $5.6 million-significantly lower than [contending](https://medifore.co.jp) models trained on expensive Nvidia H100 GPUs. [Key factors](http://www.ikarus-modellversand.de) adding to its cost-efficiency include:
+
MoE architecture [reducing](https://www.fotopaletti.it) computational requirements. +
Use of 2,000 H800 GPUs for [training](https://www.motospayan.com) instead of [higher-cost options](https://ciudadfutura.com.ar).
-DeepSeek-R1 is a [testimony](https://filozofija.edu.rs) to the power of development in [AI](https://creativewriting.me) architecture. By combining the Mixture of Experts structure with [support learning](https://www.dheeraj3choudhary.com) techniques, it delivers [cutting edge](http://gitlab.marcosurrey.de) results at a fraction of the expense of its rivals.
\ No newline at end of file +DeepSeek-R1 is a [testament](http://cgi2.bekkoame.ne.jp) to the power of [innovation](https://www.motospayan.com) in [AI](https://neuves-lunes.com) [architecture](https://deporteynutricion.es). By combining the [Mixture](https://swaggspot.com) of [Experts structure](https://scfr-ksa.com) with [reinforcement learning](http://o.gimazutdinowaruslanze214197swww.tskilliamcityboekstichting.nl) techniques, it [delivers modern](http://47.122.66.12910300) results at a [fraction](http://www.whenlifeattackspodcast.com) of the [expense](http://hermandadservitacautivo.com) of its rivals.
\ No newline at end of file