[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-saidzoda-tajik-lfm-2-6b-tool-calling":3,"topics-all":38,"news-related-125a6c5c-2930-40a5-9058-bc873b2521df":57},{"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},"125a6c5c-2930-40a5-9058-bc873b2521df","塔吉克语工具模型:CPT清零工具调用,SFT拉回64分","杜尚别 Saidzoda 实验室把 LFM2.5-2.6B-Base 适配成塔吉克语工具调用模型:CPT 注入语言时工具调用被清零,SFT 配英文回放救回,塔吉克语从 20.69 拉到 64.14;代价是 MMLU 跌到 52.5、该拒绝时说「不」的比例只剩 6.9%。","塔吉克语是塔吉克斯坦的官方语言,用西里尔字母书写,其中有六个字母连俄语都没有。开源模型的支持语言列表里几乎见不到它——直到上周五,杜尚别的 Saidzoda AI Lab 把 Saidzoda-Tajik-LFM-2.6B 挂上 arXiv:一个专门做塔吉克语工具调用的 2.6B 模型,附带 12 页技术报告。呼叫中心、政务窗口、银行系统想要的本来就不是聊天,而是把一句塔吉克语请求变成一次正确的系统调用——这正是它的定位。\n\n## 配方:两阶段 LoRA,3 张 H100 跑 47.1 小时\n\n基座是 Liquid AI 的 LFM2.5-2.6B-Base(短卷积 + 分组查询注意力的混合架构)。第一阶段 LoRA 持续预训练(CPT):1,485,319 篇清洗后的塔吉克网页文档,3 个 epoch、17,574 步,每步 96 条 4,096 token 的打包序列,3 张 H100 共 47.1 小时,训练损失从 6.615 降到 1.662。第二阶段 SFT:机器翻译的塔吉克语工具调用、61,798 行保留原文的英文工具调用(回放)、塔吉克语指令,按 40\u002F35\u002F25 配比只跑一个 epoch,全程用 axolotl 训练。\n\n## 结果:塔吉克语 20.69 到 64.14,英文不掉\n\n测试集是自建的 174 题 BFCL v3 塔吉克语翻译版(89 道单调用、27 道并行、29 道无关、29 道多轮)。严格全调用准确率:基座英语 76.55%,同样请求换成塔吉克语只剩 20.69%;适配后塔吉克语拉到 64.14,配对提升 +43.45 个百分点(95% 区间 +35.2 到 +51.7),英文 77.24 基本不动。宽松 parser(也接受基座自己的调用格式)下基座是 31.72——差距真实存在,口径一换就小了不少。\n\n## 代价:CPT 把工具调用打到 0.00,模型几乎不会说「不」\n\n最有信息量的发现在分阶段测量里:两个 CPT checkpoint 的严格得分全部是 0.00。在原始网页文本上继续预训练,把基座原生的工具调用格式整个抹掉——宽松重打分显示,一部分技能本身也跟着丢了。SFT 靠英文回放把格式和英文技能一起救回来,但账单不短:英文 MMLU 从 60.8 掉到 52.5;工具不适用时正确拒绝的比例,塔吉克语从 89.7 崩到 6.9,英文也从 62.1 掉到 31.0——它现在几乎逢请求必调工具;多轮调用仍弱(3.45 到 17.24)。模型卡直接写明:它的调用应该在执行前被检查。收益侧,塔吉克语知识基准从 28.0 涨到 35.0,Belebele 阅读理解 37.6 到 40.0、无显著变化。\n\n## 所以呢\n\n论文自述「据其所知,这是首个公开发布并附带技术报告的塔吉克语工具调用模型」——作者检索过 arXiv 与 Hugging Face Hub,2026 年 10 月未发现同类。对长尾语言社区,这套样板比模型本身值钱:数据来源、配置、算力、训练日志、每一步评测全公开,eval_results 原始文件随仓库放出;权重是 5 GB 合并 bfloat16,LFM Open License v1.0,商用受营收门槛限制,下载要先在 saidzoda.com 填表;算力来自 NVIDIA Inception 计划;模型卡致谢里也写明 Anthropic Claude 参与了脚本与文本草拟、作者核过每个数字。更通用的一课:给小模型注入新语言不是「翻译数据微调」那么便宜——CPT 会先把指令与工具格式清零,英文回放是那条护栏;而教会它调工具的另一面,是「教会它别乱调」这课还没上完。低资源语言想要 Agent,先想清楚谁来看住那个 6.9%。\n\n引用:[arXiv:2610.41576](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.41576) · [HF 模型卡](https:\u002F\u002Fhuggingface.co\u002FTohirju\u002FSaidzoda-Tajik-LFM-2.6B)\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.41576","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"fcffca1b-121d-4162-be1b-252409732f19","en","Tajik tool-calling LLM: CPT wipes the skill, SFT restores it","Tajik tool-calling on LFM2.5-2.6B: CPT zeroes the skill, SFT with English replay restores 64.14 from 20.69; MMLU falls to 52.5, refusal to 6.9%.","Tajik, the state language of Tajikistan, is written in Cyrillic with six letters that Russian does not have, and open models almost never list it among their supported languages. Last Friday, Saidzoda AI Lab in Dushanbe posted Saidzoda-Tajik-LFM-2.6B to arXiv: a 2.6B model built specifically for tool calling in Tajik, with a 12-page technical report attached. Call centres, government services and banking never wanted chat to begin with — they need a small model that turns a Tajik request into one correct call to an existing system. That is exactly the job.\n\n## The recipe: two LoRA stages, 3 H100s for 47.1 hours\n\nThe base is Liquid AI's LFM2.5-2.6B-Base, a hybrid of short convolutions and grouped-query attention. Stage one is LoRA continued pretraining (CPT): 1,485,319 cleaned Tajik web documents for 3 epochs and 17,574 optimiser steps, 96 packed sequences of 4,096 tokens each, on 3 H100 GPUs for 47.1 hours; training loss fell from 6.615 to 1.662. Stage two is SFT: machine-translated Tajik tool calls, 61,798 untranslated English tool calls kept as replay, and Tajik instructions, mixed at 40\u002F35\u002F25 for one epoch only, trained with axolotl throughout.\n\n## The result: Tajik 20.69 to 64.14, English untouched\n\nThe test set is their own 174-item Tajik translation of BFCL v3 (89 single, 27 parallel, 29 irrelevance, 29 multi-turn). Strict full-call accuracy: the base answers 76.55% of English items, but only 20.69% when the same request is written in Tajik. The adapted model lifts Tajik to 64.14, a paired gain of +43.45 points (95% interval +35.2 to +51.7), while English sits at 77.24. A lenient parser that also accepts the base's own call formats scores the base at 31.72 — the gap is real, but it narrows considerably under the lenient yardstick.\n\n## The cost: CPT zeroes tool calling to 0.00, and the model almost never says no\n\nThe most informative finding sits in the stage-by-stage measurement: every CPT checkpoint scores 0.00 in both languages. Continued pretraining on raw web text wipes out the base model's native tool-call format entirely — and, as the lenient re-scoring shows, part of the skill itself. SFT with English replay rebuilds the format and the English skill together, but the bill is long: English MMLU falls from 60.8 to 52.5; when no offered tool fits, correct refusal collapses from 89.7 to 6.9 in Tajik and from 62.1 to 31.0 in English — the model now reaches for a tool on almost every request; multi-turn calls stay weak (3.45 to 17.24). The model card says it plainly: its calls should be checked before they are executed. On the plus side, the Tajik knowledge benchmark rises from 28.0 to 35.0, while Belebele reading comprehension moves 37.6 to 40.0 with no significant change.\n\n## So what\n\nThe paper states, in its own words, that to the authors' knowledge this is the first public Tajik tool-calling model released with a technical report — the authors searched arXiv and the Hugging Face Hub in October 2026 and found none. For long-tail language communities the template is worth more than the model itself: data sources, configurations, compute, training logs and every evaluation are published, and the eval_results folder ships the original files. The weights are merged bfloat16 at about 5 GB under the LFM Open License v1.0 — commercial use is capped by a revenue threshold, and downloading requires filling a form at saidzoda.com first. Training compute came through the NVIDIA Inception programme, and the acknowledgements note that Anthropic Claude helped write scripts and draft text while the authors checked every number. The broader lesson: injecting a new language into a small model is not \"translate data and fine-tune\" — CPT zeroes the instruction and tool format first, and English replay is the guardrail that brings it back. The other half of the lesson is unfinished: teaching a model to call tools also taught it to always call. If a low-resource language wants agents, someone has to watch that 6.9%.\n\nReferences: [arXiv:2610.41576](https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.41576) · [HF model card](https:\u002F\u002Fhuggingface.co\u002FTohirju\u002FSaidzoda-Tajik-LFM-2.6B)\n","saidzoda-tajik-lfm-2-6b-tool-calling","2026-11-02T19:11:09Z","2026-11-02T19:11:14.359459Z","2026-11-02T19:11:14.359468Z",true,"agent",77,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"88d15f9c-5548-4dae-bde5-75e79c3d0a28","8K 上下文顶 1M:把长文档问题训练成 agent 接力,而不是堆窗口","trained-agentic-context-management","2026-10-01T11:32:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"5878a668-282c-4b88-b2b8-7eef40b7938c","LFM2.5-2.6B：2.5GB 内存跑本机 Agent 220 tok\u002Fs","lfm2-5-2-6b-on-device-agent","2026-08-11T00:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"15987a0e-bc06-4f21-9608-264da02d0e6c","AutoMem 让 32B 开源模型在长程任务上追平 Claude Opus 4.5","automem-stanford-32b-long-horizon","2026-07-23T12:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"64088ca6-c244-474c-93ce-cdbfe55de6d6","投机解码没用满?agent循环把加速比挖到10.87倍","specpilot-agentic-speculative-decoding","2026-11-04T13:15:17+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"90eaefa3-cbb1-440a-9c2f-97c594d0bf78","CUA-Master:教学视频喂出OSWorld 80.2","cua-master-open-world-flywheel","2026-11-03T21:08:37+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"26337a90-e049-44cb-9738-16d10a1fe84b","微软4B混合智能体:OSWorld 66%,点击还是敲代码","microsoft-mage-vl-cowork-hybrid-computer-use","2026-11-03T07:33:53+00:00"]