[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llm-fine-tuning-turns-generic-models-into-domain-tools-en":3,"article-related-llm-fine-tuning-turns-generic-models-into-domain-tools-en":30,"series-research-772c0694-0e86-465d-b676-012a2240eaf7":81},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"772c0694-0e86-465d-b676-012a2240eaf7","llm-fine-tuning-turns-generic-models-into-domain-tools-en","LLM fine-tuning turns generic models into domain tools","\u003Cp data-speakable=\"summary\">This breaks down enterprise \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> fine-tuning into a copyable workflow.\u003C\u002Fp>\u003Cp>I've been using \u003Ca href=\"\u002Ftag\u002Fllms\">LLMs\u003C\u002Fa> in production long enough to stop being impressed by demos. The first few weeks are always the same: the model sounds smart, answers fast, and makes everyone in the room nod like we solved language. Then you put it in front of real users, real policy, real product docs, and it starts doing that annoying thing where it almost gets it right. It uses the right vocabulary but the wrong rule. It sounds confident while being a little wrong in exactly the place that matters. That's the part that burns time.\u003C\u002Fp>\u003Cp>What finally clicked for me is that most teams reach for fine-tuning too early or too late. Too early, and they try to teach the model facts that should have come from retrieval. Too late, and they keep patching prompts while the model still talks in the wrong tone, misses the same format, or refuses to follow the house style. I’ve seen both. The fix is not “fine-tune everything.” The fix is knowing what fine-tuning is actually for, and what it is absolutely not for.\u003C\u002Fp>\u003Cp>The source that pushed me to write this down is \u003Ca href=\"https:\u002F\u002Faimultiple.com\u002Fllm-fine-tuning\">AI Multiple’s enterprise guide to LLM fine-tuning\u003C\u002Fa>. It makes the same basic point I keep repeating to teams: if the model lacks facts, use retrieval or a domain model; if it has the facts but not the behavior, fine-tuning starts to make sense. That distinction saves a lot of wasted training runs and a lot of expensive optimism.\u003C\u002Fp>\u003Ch2>Stop using fine-tuning as a facts bucket\u003C\u002Fh2>\u003Cblockquote>If your LLM doesn’t have access to the facts needed in your domain, either train a new LLM, switch to a domain-specific one, or use RAG to retrieve facts. If it has relevant facts but needs to answer in a different style and tone, follow certain output formats, or use certain tools, then fine-tuning is the right approach.\u003C\u002Fblockquote>\u003Cp>What this actually means is simple, even if teams keep muddying it up: fine-tuning is not your knowledge base. It is not a substitute for current information. It is not the place to dump every policy doc, every product spec, and every PDF someone found on a shared drive. If the model needs current facts, I’d rather wire in retrieval than bake stale knowledge into weights.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782569906260-hdga.png\" alt=\"LLM fine-tuning turns generic models into domain tools\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>I ran into this on a support assistant project where the team wanted the bot to answer billing questions. Their first instinct was to fine-tune on tickets. Bad idea. The tickets contained outdated policy, edge-case exceptions, and a lot of human mess. What they really needed was a retrieval layer over the current policy docs, plus a small tuning pass to make the model answer in the company’s preferred format.\u003C\u002Fp>\u003Cp>That split matters because it keeps your system maintainable. Facts change. Behavior changes slower. If you fine-tune facts, you create a retraining treadmill. If you fine-tune behavior, you get a model that sounds like your product and follows your workflow without pretending it has memorized the world.\u003C\u002Fp>\u003Cp>How to apply it:\u003C\u002Fp>\u003Cul>\u003Cli>Use \u003Ca href=\"https:\u002F\u002Fwww.pinecone.io\u002Flearn\u002Fretrieval-augmented-generation\u002F\">RAG\u003C\u002Fa> or another retrieval system when the answer depends on current documents, policies, or product data.\u003C\u002Fli>\u003Cli>Use fine-tuning when the answer is already known, but the model keeps missing format, tone, classification boundaries, or tool-use patterns.\u003C\u002Fli>\u003Cli>Use both when you need current facts and a strict response style.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That’s the cleanest mental model I know. It keeps you from paying for training when a search index would have done the job.\u003C\u002Fp>\u003Ch2>Web data is useful because it is messy in the right way\u003C\u002Fh2>\u003Cp>AI Multiple says enterprise teams increasingly use real-time web data during preparation because models are trained on fixed datasets and don’t know what changed yesterday. That point is worth underlining because it cuts against the fantasy that a pre-trained model is somehow “done.” It isn’t. It’s just frozen at a moment in time.\u003C\u002Fp>\u003Cp>What this actually means is that your training set should reflect the world your model will actually operate in. If I’m building for legal, finance, healthcare, or a niche technical domain, I want recent terminology, current workflows, and examples that look like the stuff users will ask about. Web-sourced data helps with that. Not because the web is magically clean. It’s not. It’s useful because it is current, and because it contains the language people really use.\u003C\u002Fp>\u003Cp>I’ve seen teams over-index on curated internal docs and then wonder why the model sounds like a policy manual from 2019. Fresh web data fixes part of that. It also helps expose the model to the edge cases and phrasing variations that show up in the wild. That is where hallucinations often start: the model has a gap, and it fills it with something plausible.\u003C\u002Fp>\u003Cp>AI Multiple’s example is a legal tech company collecting recent court rulings and legal blogs with web crawlers. That’s the right instinct. You are not just collecting text; you are collecting the shape of the domain. The terminology, the citations, the preferred phrasing, the things practitioners actually say.\u003C\u002Fp>\u003Cp>How to apply it:\u003C\u002Fp>\u003Cul>\u003Cli>Build a data pipeline with clear source filters before you think about training jobs.\u003C\u002Fli>\u003Cli>Prioritize recent, authoritative sources over broad scraping for the sake of volume.\u003C\u002Fli>\u003Cli>Deduplicate aggressively and remove low-signal pages, boilerplate, and near-duplicates.\u003C\u002Fli>\u003Cli>Keep a source log so you can explain where the training examples came from later.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>If you skip that cleanup, you are not training a smarter model. You are teaching it to be noisy faster.\u003C\u002Fp>\u003Ch2>Hallucination goes down when the model sees better examples\u003C\u002Fh2>\u003Cp>AI Multiple says high-quality real-world data from the web can reduce hallucination because it gives the model a more reliable source of truth. I agree with the direction, but I’d phrase it more carefully: fine-tuning doesn’t delete hallucination, it narrows the model’s confidence around the behaviors you actually want.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782569907235-6348.png\" alt=\"LLM fine-tuning turns generic models into domain tools\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>What this actually means is that the model stops improvising in the exact places where you’ve shown it the right answer pattern enough times. If your examples are consistent, the model learns that consistency. If your examples are sloppy, contradictory, or overloaded with exceptions, it learns to be sloppy in a more polished font.\u003C\u002Fp>\u003Cp>I ran into this with a document extraction workflow. The base model was decent at pulling fields, but it kept inventing missing values instead of leaving them blank. Once we fine-tuned on examples that explicitly showed “unknown,” “not provided,” and “N\u002FA” in the right contexts, the behavior changed. Not perfectly, but enough to stop the worst failures. That’s the real win: fewer confident lies in the places that matter.\u003C\u002Fp>\u003Cp>There’s also a trap here. Teams see hallucination and assume the model needs more knowledge. Sometimes it does. Often it needs a stronger refusal pattern, a better output schema, or examples that show what to do when the answer is absent. Fine-tuning is good at teaching those boundaries.\u003C\u002Fp>\u003Cp>How to apply it:\u003C\u002Fp>\u003Cul>\u003Cli>Include negative examples that show the model what not to answer.\u003C\u002Fli>\u003Cli>Teach abstention explicitly: blank, unknown, or escalate when evidence is missing.\u003C\u002Fli>\u003Cli>Use evaluation sets that measure false confidence, not just accuracy.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>If you only score exact matches, you’ll miss the dangerous part: the model sounding sure while it drifts.\u003C\u002Fp>\u003Ch2>Pick the model based on control, cost, and where the data lives\u003C\u002Fh2>\u003Cp>The source lays out a practical fork: use managed fine-tuning from a provider, or fine-tune an open-source model on-prem if you need more control over data and infrastructure. That’s the decision most enterprise teams are actually making, whether they say it out loud or not.\u003C\u002Fp>\u003Cp>What this actually means is that model choice is less about brand and more about constraints. If you need the provider to handle the training workflow, managed tuning is easier. If your data is sensitive, regulated, or expensive to move, open-source on-prem may be the safer path. If your team can’t support the compute, storage, and ops overhead, don’t pretend you can just because the model card looks friendly.\u003C\u002Fp>\u003Cp>AI Multiple mentions major providers like \u003Ca href=\"https:\u002F\u002Fcloud.google.com\u002Fvertex-ai\">Google Vertex AI\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude\">Anthropic Claude\u003C\u002Fa> via \u003Ca href=\"https:\u002F\u002Faws.amazon.com\u002Fbedrock\u002F\">Amazon Bedrock\u003C\u002Fa>, and the changing status of \u003Ca href=\"\u002Ftag\u002Fopenai\">OpenAI\u003C\u002Fa> fine-tuning. That part matters because the tooling landscape is not static. If you are planning a long-lived enterprise workflow, you need to know which provider will actually support the tuning path you’re betting on.\u003C\u002Fp>\u003Cp>I’ve had teams pick a model because it benchmarked well, then get stuck when the fine-tuning path didn’t fit their governance or deployment needs. Performance is only one axis. The other axis is operational sanity.\u003C\u002Fp>\u003Cp>How to apply it:\u003C\u002Fp>\u003Cul>\u003Cli>Choose managed tuning when speed matters more than full control.\u003C\u002Fli>\u003Cli>Choose open-source when data locality, auditability, or custom infra matters more.\u003C\u002Fli>\u003Cli>Check inference cost after tuning, not just training cost. That’s where budgets quietly die.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>And yes, you should test the deployment path before you train anything. I’ve learned that one the hard way.\u003C\u002Fp>\u003Ch2>Data prep is the part everyone underestimates\u003C\u002Fh2>\u003Cp>AI Multiple breaks the process into dataset prep, model choice, and tuning, which is about right. The part people skip is that the dataset is the product. The training run is just the ceremony.\u003C\u002Fp>\u003Cp>What this actually means is that your examples need to be shaped for the task, not just collected. If you are tuning for support replies, then each example should show the input, the desired output, and the exact style you want. If you are tuning for classification, label consistency matters more than prose quality. If you are tuning for tool use, the sequence of actions matters more than eloquence.\u003C\u002Fp>\u003Cp>I’ve seen datasets where every example was technically relevant and practically useless. Too much variation in tone. Too many labels with fuzzy boundaries. Too many examples that mixed multiple tasks in one record. The model learned the confusion faithfully. That is not a bug in the model; that’s bad supervision.\u003C\u002Fp>\u003Cp>Here’s the rough prep checklist I use before I let anyone near a training job:\u003C\u002Fp>\u003Cul>\u003Cli>Normalize labels and response formats.\u003C\u002Fli>\u003Cli>Remove duplicates and near-duplicates.\u003C\u002Fli>\u003Cli>Split by task, not just by source.\u003C\u002Fli>\u003Cli>Hold out a test set that reflects production failure cases.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That last one matters. If your evaluation set is too clean, you will ship a model that looks good in review and then falls apart in the wild. I want ugly examples in evaluation. The model should prove itself where users are most likely to break it.\u003C\u002Fp>\u003Ch2>Continuous tuning only makes sense if the world keeps changing\u003C\u002Fh2>\u003Cp>AI Multiple mentions tuning checkpoints and continuous tuning in provider tooling, which is a good reminder that fine-tuning doesn’t have to be a one-shot event. Sometimes the model needs to keep learning as the domain shifts.\u003C\u002Fp>\u003Cp>What this actually means is that continuous tuning is useful when your product or industry changes often enough that static examples go stale. New product lines, new regulations, new terminology, new workflows. If the domain is moving, your model should not be frozen in place for a year and then “updated” in a panic.\u003C\u002Fp>\u003Cp>But I’m cautious here. Continuous tuning sounds elegant until you realize you’ve created a moving target with no stable baseline. If you do this, you need versioning, rollback, and clear evaluation gates. Otherwise every update becomes a mystery box and nobody knows which training pass caused the regression.\u003C\u002Fp>\u003Cp>I like checkpointing because it gives me a way back. If a new tuning pass hurts one important slice, I want the ability to compare against the previous state and stop the bleed. That is boring engineering, which is exactly why it works.\u003C\u002Fp>\u003Cp>How to apply it:\u003C\u002Fp>\u003Cul>\u003Cli>Version datasets, prompts, and model checkpoints together.\u003C\u002Fli>\u003Cli>Run slice-based evaluations after every tuning pass.\u003C\u002Fli>\u003Cli>Keep a rollback path for the last good model.\u003C\u002Fli>\u003C\u002Ful>\u003Cp>That’s not glamorous. It’s just what keeps an enterprise workflow from turning into an expensive guessing game.\u003C\u002Fp>\u003Ch2>The template you can copy\u003C\u002Fh2>\u003Cpre>\u003Ccode># Enterprise LLM Fine-Tuning Playbook\n\n## 1) Decide whether fine-tuning is even the right tool\nUse fine-tuning only when the model already has the facts, but it still misses:\n- tone\n- format\n- task boundaries\n- tool-use patterns\n- refusal behavior\n\nIf the model needs current facts, use retrieval first.\nIf the model needs domain facts that change often, consider a domain-specific model or RAG.\n\n## 2) Define the task in one sentence\nWrite one sentence that says exactly what the tuned model should do.\nExample:\n\"Classify support tickets into one of 12 categories and respond in our house style.\"\n\n## 3) Build the dataset\nCollect examples that are:\n- current\n- domain-relevant\n- label-consistent\n- close to production inputs\n\nFor each example, store:\n- input\n- ideal output\n- source\n- label\u002Fversion\n- notes about edge cases\n\n## 4) Clean the data\nBefore training:\n- remove duplicates\n- normalize labels\n- strip boilerplate\n- separate tasks\n- exclude low-quality or contradictory examples\n\n## 5) Choose the model path\nManaged tuning if you want speed and provider tooling.\nOpen-source on-prem if you need data locality and control.\n\n## 6) Pick the tuning method\n- Supervised fine-tuning for direct input-output behavior\n- Preference tuning for ranking better responses\n- Continuous tuning only if the domain changes often\n\n## 7) Evaluate before shipping\nMeasure:\n- accuracy on core tasks\n- refusal quality\n- hallucination rate\n- format compliance\n- performance on ugly edge cases\n\n## 8) Lock the rollout\nShip with:\n- model version\n- dataset version\n- evaluation report\n- rollback plan\n- owner for retraining decisions\n\n## 9) Keep the model honest\nIf facts change, refresh retrieval.\nIf behavior drifts, retrain.\nIf the task changes, redefine the task first.\n\n## Example training record\n{\n  \"input\": \"Customer asks whether refund is available after 45 days.\",\n  \"output\": \"Refunds are available within 30 days of purchase. After that window, we can offer store credit if the item meets policy requirements.\",\n  \"label\": \"billing_policy\",\n  \"source\": \"policy-doc-v12\",\n  \"notes\": \"Explicitly refuse inventing exceptions\"\n}\n\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>This is the version I wish more teams started with. It keeps the model work grounded in actual behavior, not vibes. And it makes the next person on your team less likely to inherit a mystery machine.\u003C\u002Fp>\u003Cp>If you want the short version, here it is: use retrieval for facts, fine-tuning for behavior, and checkpoints for sanity. That’s the whole trick.\u003C\u002Fp>\u003Cp>Source attribution: I broke this down from \u003Ca href=\"https:\u002F\u002Faimultiple.com\u002Fllm-fine-tuning\">AI Multiple’s LLM Fine-Tuning Guide for Enterprises\u003C\u002Fa>. The structure, examples, and copy-ready workflow here are my own interpretation of that source, not a verbatim rewrite.\u003C\u002Fp>","A practical breakdown of enterprise LLM fine-tuning, from data prep to model choice, plus a copy-ready template.","aimultiple.com","https:\u002F\u002Faimultiple.com\u002Fllm-fine-tuning",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782569906260-hdga.png","research","en","7c4c30b3-b2a8-48a7-b2ea-96c40c16ae19",[17,18,19,20,21],"llm fine-tuning","enterprise ai","rag","web data","hallucination",[23,24,25],"Fine-tuning is for behavior, not fresh facts.","Web data helps when the domain changes faster than your model.","Dataset quality matters more than the training run itself.",0,"2026-06-27T14:17:57.190952+00:00","2026-06-27T14:17:57.178+00:00","3103988e-c4fe-45e3-98ab-846500c9d507",{"tags":31,"relatedLang":40,"relatedPosts":44},[32,34,37],{"name":33,"slug":19},"RAG",{"name":35,"slug":36},"enterprise AI","enterprise-ai",{"name":38,"slug":39},"LLM fine-tuning","llm-fine-tuning",{"id":15,"slug":41,"title":42,"language":43},"llm-fine-tuning-turns-generic-models-into-domain-tools-zh","LLM 微調把通用模型變專用工具","zh",[45,51,57,63,69,75],{"id":46,"slug":47,"title":48,"cover_image":49,"image_url":49,"created_at":50,"category":13},"25aef6a0-efaa-459c-bca4-77f0d462b792","rust-learners-need-permission-to-clone-first-en","Rust learners need permission to clone first, optimize later","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782552763890-fem3.png","2026-06-27T09:32:21.788692+00:00",{"id":52,"slug":53,"title":54,"cover_image":55,"image_url":55,"created_at":56,"category":13},"567f2a82-494e-493a-9d43-00dfbc8a7bfd","mistral-ocr-4-document-ai-structure-en","Mistral OCR 4 brings structure to document AI","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782468180808-ulcg.png","2026-06-26T10:02:37.910976+00:00",{"id":58,"slug":59,"title":60,"cover_image":61,"image_url":61,"created_at":62,"category":13},"de74bbd4-e3b6-407a-998b-b38c4170b586","autoregressive-boltzmann-generators-ditch-flows-en","Autoregressive Boltzmann Generators ditch flows","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782455575877-62qe.png","2026-06-26T06:32:30.585573+00:00",{"id":64,"slug":65,"title":66,"cover_image":67,"image_url":67,"created_at":68,"category":13},"c05899fc-dd62-4fad-a249-9748376c1ef2","river-llm-reinforcement-learning-without-answers-en","RiVER trains LLMs without ground-truth answers","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782454678234-6mk1.png","2026-06-26T06:17:27.491779+00:00",{"id":70,"slug":71,"title":72,"cover_image":73,"image_url":73,"created_at":74,"category":13},"696a4c45-6c7b-4a78-a947-2dee1ddc4a58","danceopd-on-policy-generative-field-distillation-en","DanceOPD distills image-editing skills into one model","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782453779169-rakb.png","2026-06-26T06:02:33.604728+00:00",{"id":76,"slug":77,"title":78,"cover_image":79,"image_url":79,"created_at":80,"category":13},"2ae923b8-38e0-402a-937e-8e085f6a022d","microsoft-ai-team-collaboration-cfp-2026-en","Microsoft funds AI research on team collaboration","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782415977730-aqdi.png","2026-06-25T19:32:33.635454+00:00",[82,87,92,97,102,107,112,117,122,127],{"id":83,"slug":84,"title":85,"created_at":86},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":128,"slug":129,"title":130,"created_at":131},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]