[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-human-vs-synthetic-post-training-budget":3,"topics-all":38,"news-related-75867877-77f5-40dd-a467-115164eb4071":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"75867877-77f5-40dd-a467-115164eb4071","合成数据三四百美元到顶，人工1.3万才追平","后训练的偏好数据，该雇人标注还是让模型自己生成？论文把两边预算钉成美元对账，三个7-8B基座实测：合成偏好数据每条0.0102美元，三四百美元就逼近胜率顶；人工每条约1.3美元，1.3万美元档才在助人任务上追平，摘要任务始终没追上；两种混着用两头落空，人工源SHP更是全线跌破未训练基线。","后训练阶段绕不开一笔账：SFT 之后还想做偏好对齐，数据从哪来？雇人标偏好，贵，但一直被当成金标准；用大模型自己造，便宜，又总被怀疑自欺欺人。过去比较这两条路线，几乎都固定样本数——各拿一万条比效果。Shourik Banerjee 等人 11 月 12 日发布在 [arXiv:2611.16895](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16895) 的论文换了个问法：把预算钉成美元，两边各自采购，谁的模型更强？\n\n## 把钱钉死再比\n\n三个 7B 级开源基座：Olmo-3-7B、Llama-3.1-8B、Qwen-2.5-7B，统一走 SFT 加 DPO 加 LoRA。人工侧用四个公开偏好数据集：HelpSteer3、SHP、Anthropic HH、TL;DR。合成侧自己造两套，结构分别对齐 Anthropic HH 与 TL;DR：生成器是 Llama-3.1-Tulu-3-8B-SFT，靠 Persona Hub 人设采样保多样性，再让 Qwen-3-32B 当裁判、每对回答投 32 票取多数。\n\n成本按公开牌价折算：合成偏好数据每条 0.0102 美元，摘要任务每条 0.0194 美元；人工标注按复现口径估每条约 1.3 美元——助人场景两类数据单价差约 130 倍。预算档统一设 100、200、300、400、1000 美元，人工侧额外加一个 1.3 万美元档（一万条）。评测用 2048 条 held-out 测试集、gpt-oss-20b 当裁判、候选位置随机化；论文还专门验证了裁判的可信度：它与两位人类标注员的一致率，和这两位人类彼此之间的一致率相当。\n\n## 合成先到顶，人工要用十倍以上的钱\n\n三个基座规律一致：合成数据前几百美元涨得飞快，随后饱和，多数提升发生在前几百美元内——天花板由数据质量决定，而不是预算。人工数据正好相反：同样几百美元只买得起 76 到 769 条样本，模型几乎原地不动；要等到 1.3 万美元档才有竞争力，比合成侧顶格的 1000 美元档还贵十倍以上，且只有 Llama-3.1-8B 和 Qwen-2.5-7B 超过合成峰值，Olmo-3-7B 仍没追上。助人任务上，在双方都训练过的每个预算点上，合成数据都追平或超过所有人工源。\n\n摘要任务更残酷：TL;DR 人工数据到 1.3 万美元档仍低于合成天花板，任何测试过的预算下都没追上。Qwen-2.5-7B 还走出一条先跌后升的曲线——基线胜率 0.41，小数据量微调先跌到 0.20，喂到 12000 条、约 1.56 万美元才回到 0.53，仍不及合成。\n\n## 两个反直觉发现\n\n第一，混着用更差。1.3 万美元预算下，从 95\u002F5 扫到 99\u002F1 的每种人工\u002F合成配比，胜率几乎全线低于两个纯源中较好的那个，有时连两个纯源都不如——Llama-3.1-8B 摘要任务混合配比落在 0.52 到 0.56，纯合成天花板 0.72，纯人工 0.59。根子在成本结构：130:1 的单价差下，90\u002F10 的预算分账换成样本数已经是 14.4:1；论文猜测，混合稀释了本可独自表现更好的那一方。\n\n第二，人工数据也会翻车。SHP 的偏好是从 Reddit 分数反推的，不是标注员按指南比出来的；用它训练，三个基座全部跌破未训练基线，域限定加分数过滤也救不回来。论文的提醒相当扎心：偏好是怎么产生的、回答格式合不匹配，至少和「是不是人标的」同样重要——人工源也该先审计，再当金标准。\n\n## 边界与启发\n\n论文把边界划得很清楚：只测了 7-8B 基座加 LoRA 微调，没用全参；全流程只用开源模型，结论未必迁移到闭源前沿管线；合成数据的质量上限被 8B 生成器封死，换更强的生成器，成本-性能曲线可能改写。1.3 美元的人工单价也是复现成本估计，不是当年账单。\n\n对实践者的启示很直接：预算有限，先用合成数据快速到顶；人工预算留给值得为更高天花板买单的场景，而且花钱之前先审源。论文还附赠一个免费的预警信号——过早自终止率：Olmo-3-7B 在 1000 美元档胜率从 0.876 掉到 0.794 时，句子没写完的比率从 5.5% 升到 17.5%，可以当低成本的检查点，但论文提醒它只是检查信号，不是停止规则。\n\n所以呢——如果明年数据预算砍掉一个零，先砍哪头？这篇论文的答案大概是：先砍那个没被审计过的人工数据源。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16895","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9936766d-3ab1-494b-93bc-5b97ec9b8674","en","Synthetic Data Nears Its Ceiling by $400; Human Needs $13k","Synthetic pairs at $0.0102 each near the ceiling within a few hundred dollars; human labels catch up only at $13k, and never on summarization.","Post-training comes down to a ledger: after SFT you still want preference alignment, so where does the data come from? Hiring human annotators is expensive but long treated as the gold standard; having a model generate the data is cheap but suspected of self-deception. Past comparisons almost always held sample count fixed — take ten thousand pairs from each side and compare. A paper posted on [arXiv:2611.16895](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16895) on November 12 by Shourik Banerjee and colleagues flips the question: pin the budget in dollars, let each side shop on its own, and see whose model ends up stronger.\n\n## Pin the dollars, then compare\n\nThree open 7B-class base models — Olmo-3-7B, Llama-3.1-8B, Qwen-2.5-7B — all fine-tuned with SFT, then DPO, with LoRA. The human side draws on four public preference datasets: HelpSteer3, SHP, Anthropic HH, and TL;DR. The synthetic side builds two new sets structurally mirroring Anthropic HH and TL;DR: Llama-3.1-Tulu-3-8B-SFT generates the conversations, Persona Hub personas keep them diverse, and a Qwen-3-32B judge casts 32 votes per response pair, majority wins.\n\nCosts are converted at public per-token rates: $0.0102 per synthetic preference pair ($0.0194 for summarization), while human annotation is estimated at about $1.30 per sample in replication terms — roughly a 130:1 unit-price gap on helpfulness. Shared budget tiers run $100, $200, $300, $400, and $1,000, with an extra $13k tier (10k samples) for human data. Evaluation uses a held-out 2,048-sample test set, a gpt-oss-20b judge, randomized candidate order; the paper also validates the judge — its agreement with two human annotators is on par with the agreement between those two humans.\n\n## Synthetic peaks early; human needs ten times more\n\nThe pattern holds across all three bases: synthetic data climbs fast within the first few hundred dollars, then saturates — the ceiling is set by data quality, not budget. Human data behaves oppositely: the same few hundred dollars buy only 76 to 769 examples, and trained models stay near the untrained baseline. Human data becomes competitive only at the $13k tier, more than ten times the most expensive synthetic tier, and even there it surpasses the synthetic peak only for Llama-3.1-8B and Qwen-2.5-7B — Olmo-3-7B falls short. On helpfulness, at every budget where both sources were trained, synthetic data matches or exceeds every human source.\n\nSummarization is harsher: human TL;DR data remains below the synthetic ceiling at the $13k tier and never closes the gap at any tested budget. Qwen-2.5-7B traces a dip-then-recover curve — baseline win rate 0.41, dropping to 0.20 under small fine-tunes, recovering to 0.53 only at 12,000 samples (about $15,600), still below synthetic.\n\n## Two counterintuitive findings\n\nFirst, mixing makes it worse. At a fixed $13k, every human\u002Fsynthetic split from 95\u002F5 to 99\u002F1 lands below the better of the two pure sources, sometimes below both — on Llama-3.1-8B summarization, mixed splits score 0.52 to 0.56 against a pure-synthetic ceiling of 0.72 and pure-human 0.59. The root is cost structure: at a 130:1 price gap, a 90\u002F10 dollar split already means a 14.4:1 sample ratio; the authors suspect the mixture dilutes whichever source would have performed better alone.\n\nSecond, human data can fail too. SHP preferences are inferred from Reddit scores rather than produced by annotators following guidelines; training on it drops all three bases below their untrained baselines, and domain restriction plus score filtering do not rescue it. The point is blunt: how a preference was produced, and whether the response format matches, matter at least as much as whether a human was involved — audit a human source before treating it as gold.\n\n## Boundaries and takeaways\n\nThe authors are explicit about limits: only 7-8B bases with LoRA, no full-parameter tuning; an all-open-weight pipeline whose conclusions may not transfer to proprietary frontier stacks; and a synthetic ceiling sealed by an 8B generator — a stronger generator could reshape the cost-performance curve. The $1.30 human price is a replication estimate, not a historical bill.\n\nFor practitioners the implication is direct: on a tight budget, start with synthetic data and reach the ceiling fast; reserve human annotation for settings where the higher ceiling justifies the cost, and vet the source before paying. The paper also offers a free early-warning signal — the premature self-termination rate: when Olmo-3-7B helpfulness win rate fell from 0.876 to 0.794 at the $1,000 tier, the share of responses ending mid-sentence rose from 5.5% to 17.5%. Treat it as a cheap checkpoint, the authors caution, not a stopping rule.\n\nSo — if your data budget loses a zero next year, which side do you cut first? The likely answer from this paper: cut the un-audited human source first.","human-vs-synthetic-post-training-budget","2026-11-13T13:20:00Z","2026-11-13T13:10:37.488339Z","2026-11-13T13:10:37.488349Z",true,"agent",231,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"261b6911-0300-4d6b-86dc-a64475e4aa5a","Delta N-gram:给大模型记忆打可拆卸补丁","delta-ngram-forward-only-memory-writes","2026-10-25T23:50:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"34d309c3-5757-469a-85ff-b524c8865037","Queen 模型边下棋边解说:七轮自蒸馏涨 900 Elo","queen-chess-language-model-self-distillation","2026-10-05T23:04:34+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"c361090b-291c-4357-9bd3-54845025f29d","CaRE-KD 让蒸馏学会看「教师脸色」:NeurIPS 2026 选中 IIT Delhi","care-kd-confidence-distillation-neurips-2609-36734","2026-10-05T03:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"2dbc7c0f-083a-47d4-ba9a-8a7d66b22ae0","亚马逊八阶段配方:后训练让 GLM-4.5-Air 反超官方版","amazon-rufus-air-post-training-recipe","2026-09-25T23:12:24+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"054e060c-e182-42a9-b3ed-229feb8ac0ac","2026 年的蒸馏长什么样:Hugging Face 拆解前沿模型三大范式","2026-distillation-three-paradigms","2026-07-09T12:00:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"ab2d6e9e-8890-4ae7-b6ca-8febc831a279","HPLT MultiSynt\u002FMT：4.8 万亿 token 多语种数据集","hplt-multisynt-multilingual-dataset","2026-07-07T16:01:00+00:00"]