[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F1B0ZTP2HKYPVB93GT0G85":3,"skill-same:01M7F1B0ZTP2HKYPVB93GT0G85":56},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7F1B0ZTP2HKYPVB93GT0G85","lora-qlora-recipes",null,"Configure LoRA and QLoRA supervised fine-tuning with current best-practice hyperparameters. Use when writing or reviewing a LoRA\u002FQLoRA training configuration, choosing rank\u002Falpha\u002Ftarget modules, or deciding between LoRA, QLoRA, and full fine-tuning.","shotcowboystyle\u002Fag-skills","shotcowboystyle","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F62471?v=4","01M7EZTCFCZJ6K3MYW8BPM4X6R","Curtis Blanton","01M7EZTCG1S5M04J90A931X9BX",0,"MIT","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Flora-qlora-recipes\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: lora-qlora-recipes\ndescription: Configure LoRA and QLoRA supervised fine-tuning with current best-practice hyperparameters. Use when writing or reviewing a LoRA\u002FQLoRA training configuration, choosing rank\u002Falpha\u002Ftarget modules, or deciding between LoRA, QLoRA, and full fine-tuning.\n---\n\n# LoRA & QLoRA Recipes\n\nThis skill assumes the routing decision already\nhappened — `finetuning-method-selection` should\nhave already pointed here because the data shape\nis demonstrations (SFT), not preference pairs or\na verifiable reward signal. What follows is the\ncurrent best-practice recipe for configuring the\nadapter itself: which modules to target, how to\nsize rank and alpha, what learning rate to use,\nand when QLoRA buys real headroom versus when it\njust adds risk. Dataset preparation and quality\nchecks are a separate concern — see\n`dataset-curation`.\n\n**Input:** a routing decision (SFT via LoRA\u002F\nQLoRA) plus a target size class.\n**Output format:** a validated adapter config —\nthe kwarg values below, not free-form advice —\nthat `llm-finetuning-training-engineer` consumes\ndirectly when it generates a runnable script.\n\n## The Reference Recipe\n\nThe reference recipe is \"LoRA Without Regret\"\n(Thinking Machines\u002FSchulman, 2025-09), now the\nsettled convention for LoRA\u002FQLoRA SFT.\n\n### Target Modules\n\nTarget **all-linear** modules, not just\nattention:\n\n```python\ntarget_modules = [\n    \"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",   # attention\n    \"gate_proj\", \"up_proj\", \"down_proj\",      # MLP — matters most\n]\n```\n\nThe MLP layers (`gate_proj`, `up_proj`,\n`down_proj`) matter most — attention-only\ntargeting was the older, weaker convention.\nDropping modules to save memory is a Failure\nMode below, not a valid optimization.\n\n### Alpha and Learning Rate\n\n- **`lora_alpha = 2 * r`** is the settled\n  convention (NeurIPS 2025 \"intruder dimensions\"\n  result). Don't hand-tune alpha independently of\n  rank — derive it from rank every time.\n- **LoRA learning rate ≈ 10x the equivalent\n  full-fine-tune LR.** For QLoRA specifically,\n  **2e-4** is the standard starting point. The rule\n  carries over to DPO, ORPO, SimPO, KTO and GRPO\n  adapters. Full hyperparameter tables, the\n  per-adapter-type table for those methods, and\n  worked examples:\n  `references\u002Fhyperparameters.md`.\n\n### Rank by Task\n\nRank is task-shaped, not a single global default:\n\n| Task | Rank |\n|---|---|\n| RL (GRPO\u002FRLVR adapters) | 1–32 |\n| General default | 16–32 |\n| SFT at scale | up to ~256 |\n\nHigher rank isn't automatically better — it\nraises capacity to memorize as fast as it raises\ncapacity to generalize. Start at the row matching\nthe task, and only move up a row if the lower\nrank measurably underfits on held-out eval, not\nas a default hedge.\n\n### Effective Batch Size\n\nKeep **effective batch size under 32**. This\nrecipe was validated at that scale — pushing\neffective batch higher is an untested\nextrapolation, not a free throughput win.\n\n## Unsloth Defaults\n\nUnsloth is the reference implementation this\nplugin assumes as the default fast path — except\nfor messages-shaped conversational SFT with\n`assistant_only_loss=True`, where Unsloth\n2026.7.x's compiled trainer has no messages-shaped\npath at all and the plain-TRL escape hatch\n(`references\u002Funsloth-trl-mapping.md`) is the\ndefault for that combination, not a rare-regression\nfallback. Its out-of-the-box defaults, and why\neach one is set that way:\n\n- **`lora_dropout=0`** — the optimized kernel\n  path assumes zero dropout; setting a nonzero\n  value forfeits the fused-kernel speedup.\n- **`bias=\"none\"`** — bias terms add adapter\n  parameters for negligible quality gain at this\n  rank range.\n- **`use_gradient_checkpointing=\"unsloth\"`** —\n  Unsloth's checkpointing variant, not vanilla HF\n  checkpointing; saves roughly **30% VRAM** over\n  no checkpointing.\n- **`optim=\"adamw_8bit\"`** — 8-bit AdamW cuts\n  optimizer-state memory with negligible quality\n  impact at LoRA\u002FQLoRA adapter scale.\n- **`random_state`** fixed — pins LoRA\n  initialization for reproducibility across runs;\n  treat it like any other seed, not a tunable.\n\nThese show up together on the `get_peft_model`\ncall:\n\n```python\nmodel = FastLanguageModel.get_peft_model(\n    model,\n    r=32,\n    target_modules=target_modules,\n    lora_alpha=64,               # 2 * r\n    lora_dropout=0,\n    bias=\"none\",\n    use_gradient_checkpointing=\"unsloth\",\n    random_state=3407,\n)\n```\n\nExact kwarg names and their plain-TRL\u002FPEFT\nequivalents, plus a full worked config including\n`SFTConfig`: `references\u002Funsloth-trl-mapping.md`\nand `references\u002Fhyperparameters.md`.\n\n## LoRA vs QLoRA vs Full FT\n\n| Situation | Default choice |\n|---|---|\n| Adapting behavior on demonstrations | LoRA |\n| Base model doesn't fit in bf16 at target rank | QLoRA |\n| Injecting dense new domain knowledge | Full FT (see `finetuning-method-selection`) |\n| Unsure which one | LoRA — upgrade to QLoRA only if memory forces it |\n\n- **QLoRA** = NF4-quantized frozen base weights +\n  BF16 adapters. This is what makes a 65B-class\n  model trainable on 48GB — the quantized base\n  is the memory win, not the adapter itself.\n- **Full fine-tuning is not a default.** Reserve\n  it for dense knowledge injection where the goal\n  is changing what the model knows at the weight\n  level, not adapting a behavior. For everything\n  else in this skill's scope, LoRA or QLoRA is\n  the starting assumption.\n- **On DGX Spark, QLoRA can OOM before an\n  equivalent bf16 LoRA run would**, even though\n  QLoRA's steady-state footprint is smaller —\n  bitsandbytes dequantization buffers are\n  transient CUDA-side allocations that spike\n  during load. A QLoRA OOM is not proof the model\n  doesn't fit; the `dgx-spark-ops` plugin's\n  `spark-memory-thermal-ops` skill covers the\n  full OOM remediation ladder (bf16 LoRA is the\n  next thing to try, not a further QLoRA\n  shrink).\n\n## Failure Modes\n\n- **fp16 divergence on non-BF16 GPUs.** Training\n  in fp16 on hardware that doesn't have solid\n  BF16 support is a known source of loss spikes\n  and silent divergence. Force `bf16=True`\n  wherever the hardware supports it; don't fall\n  back to fp16 as if it were equivalent. Check\n  hardware support before picking a dtype:\n\n  ```bash\n  python -c \"import torch; print(torch.cuda.is_bf16_supported())\"\n  ```\n\n- **Rank too high on a small dataset overfits.**\n  A rank picked for \"SFT at scale\" (up to ~256)\n  on a dataset that doesn't have scale behind it\n  memorizes rather than generalizes. Match rank\n  to the Rank by Task table above, not to the\n  largest number available.\n- **Removing target modules to save memory costs\n  quality for negligible savings.** The adapter\n  parameters on `gate_proj`\u002F`up_proj`\u002F`down_proj`\n  are a small fraction of total model size — cutting\n  them barely moves memory but measurably hurts\n  quality. If memory is tight, move to QLoRA or\n  reduce rank\u002Fbatch\u002Fpack length before trimming\n  target modules.\n\nAll three failure modes share a pattern: they\nlook like a training-loop bug (loss spikes,\nplateaus, memorization) but are actually a\nconfig choice that contradicts the reference\nrecipe above. Check configuration against this\nskill before debugging the training loop itself.\n\n## References\n\n- `references\u002Fhyperparameters.md` — full rank\u002F\n  alpha\u002FLR tables by task type, rsLoRA notes,\n  batch\u002Fpacking interactions, and a complete\n  worked Unsloth config block.\n- `references\u002Funsloth-trl-mapping.md` — every\n  Unsloth kwarg mapped to its TRL\u002FPEFT\n  equivalent, current TRL API notes, and the\n  escape-hatch rule for when to drop back to\n  plain TRL.\n\nRelated skills: `finetuning-method-selection`\nroutes here; `dataset-curation` covers the data\nside this skill doesn't; `llm-finetuning-training-engineer`\nis the downstream consumer of the config this\nskill produces.\n","# LoRA & QLoRA Recipes\n\nThis skill assumes the routing decision already\nhappened — `finetuning-method-selection` should\nhave already pointed here because the data shape\nis demonstrations (SFT), not preference pairs or\na verifiable reward signal. What follows is the\ncurrent best-practice recipe for configuring the\nadapter itself: which modules to target, how to\nsize rank and alpha, what learning rate to use,\nand when QLoRA buys real headroom versus when it\njust adds risk. Dataset preparation and quality\nchecks are a separate concern — see\n`dataset-curation`.\n\n**Input:** a routing decision (SFT via LoRA\u002F\nQLoRA) plus a target size class.\n**Output format:** a validated adapter config —\nthe kwarg values below, not free-form advice —\nthat `llm-finetuning-training-engineer` consumes\ndirectly when it generates a runnable script.\n\n## The Reference Recipe\n\nThe reference recipe is \"LoRA Without Regret\"\n(Thinking Machines\u002FSchulman, 2025-09), now the\nsettled convention for LoRA\u002FQLoRA SFT.\n\n### Target Modules\n\nTarget **all-linear** modules, not just\nattention:\n\n```python\ntarget_modules = [\n    \"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",   # attention\n    \"gate_proj\", \"up_proj\", \"down_proj\",      # MLP — matters most\n]\n```\n\nThe MLP layers (`gate_proj`, `up_proj`,\n`down_proj`) matter most — attention-only\ntargeting was the older, weaker convention.\nDropping modules to save memory is a Failure\nMode below, not a valid optimization.\n\n### Alpha and Learning Rate\n\n- **`lora_alpha = 2 * r`** is the settled\n  convention (NeurIPS 2025 \"intruder dimensions\"\n  result). Don't hand-tune alpha independently of\n  rank — derive it from rank every time.\n- **LoRA learning rate ≈ 10x the equivalent\n  full-fine-tune LR.** For QLoRA specifically,\n  **2e-4** is the standard starting point. The rule\n  carries over to DPO, ORPO, SimPO, KTO and GRPO\n  adapters. Full hyperparameter tables, the\n  per-adapter-type table for those methods, and\n  worked examples:\n  `references\u002Fhyperparameters.md`.\n\n### Rank by Task\n\nRank is task-shaped, not a single global default:\n\n| Task | Rank |\n|---|---|\n| RL (GRPO\u002FRLVR adapters) | 1–32 |\n| General default | 16–32 |\n| SFT at scale | up to ~256 |\n\nHigher rank isn't automatically better — it\nraises capacity to memorize as fast as it raises\ncapacity to generalize. Start at the row matching\nthe task, and only move up a row if the lower\nrank measurably underfits on held-out eval, not\nas a default hedge.\n\n### Effective Batch Size\n\nKeep **effective batch size under 32**. This\nrecipe was validated at that scale — pushing\neffective batch higher is an untested\nextrapolation, not a free throughput win.\n\n## Unsloth Defaults\n\nUnsloth is the reference implementation this\nplugin assumes as the default fast path — except\nfor messages-shaped conversational SFT with\n`assistant_only_loss=True`, where Unsloth\n2026.7.x's compiled trainer has no messages-shaped\npath at all and the plain-TRL escape hatch\n(`references\u002Funsloth-trl-mapping.md`) is the\ndefault for that combination, not a rare-regression\nfallback. Its out-of-the-box defaults, and why\neach one is set that way:\n\n- **`lora_dropout=0`** — the optimized kernel\n  path assumes zero dropout; setting a nonzero\n  value forfeits the fused-kernel speedup.\n- **`bias=\"none\"`** — bias terms add adapter\n  parameters for negligible quality gain at this\n  rank range.\n- **`use_gradient_checkpointing=\"unsloth\"`** —\n  Unsloth's checkpointing variant, not vanilla HF\n  checkpointing; saves roughly **30% VRAM** over\n  no checkpointing.\n- **`optim=\"adamw_8bit\"`** — 8-bit AdamW cuts\n  optimizer-state memory with negligible quality\n  impact at LoRA\u002FQLoRA adapter scale.\n- **`random_state`** fixed — pins LoRA\n  initialization for reproducibility across runs;\n  treat it like any other seed, not a tunable.\n\nThese show up together on the `get_peft_model`\ncall:\n\n```python\nmodel = FastLanguageModel.get_peft_model(\n    model,\n    r=32,\n    target_modules=target_modules,\n    lora_alpha=64,               # 2 * r\n    lora_dropout=0,\n    bias=\"none\",\n    use_gradient_checkpointing=\"unsloth\",\n    random_state=3407,\n)\n```\n\nExact kwarg names and their plain-TRL\u002FPEFT\nequivalents, plus a full worked config including\n`SFTConfig`: `references\u002Funsloth-trl-mapping.md`\nand `references\u002Fhyperparameters.md`.\n\n## LoRA vs QLoRA vs Full FT\n\n| Situation | Default choice |\n|---|---|\n| Adapting behavior on demonstrations | LoRA |\n| Base model doesn't fit in bf16 at target rank | QLoRA |\n| Injecting dense new domain knowledge | Full FT (see `finetuning-method-selection`) |\n| Unsure which one | LoRA — upgrade to QLoRA only if memory forces it |\n\n- **QLoRA** = NF4-quantized frozen base weights +\n  BF16 adapters. This is what makes a 65B-class\n  model trainable on 48GB — the quantized base\n  is the memory win, not the adapter itself.\n- **Full fine-tuning is not a default.** Reserve\n  it for dense knowledge injection where the goal\n  is changing what the model knows at the weight\n  level, not adapting a behavior. For everything\n  else in this skill's scope, LoRA or QLoRA is\n  the starting assumption.\n- **On DGX Spark, QLoRA can OOM before an\n  equivalent bf16 LoRA run would**, even though\n  QLoRA's steady-state footprint is smaller —\n  bitsandbytes dequantization buffers are\n  transient CUDA-side allocations that spike\n  during load. A QLoRA OOM is not proof the model\n  doesn't fit; the `dgx-spark-ops` plugin's\n  `spark-memory-thermal-ops` skill covers the\n  full OOM remediation ladder (bf16 LoRA is the\n  next thing to try, not a further QLoRA\n  shrink).\n\n## Failure Modes\n\n- **fp16 divergence on non-BF16 GPUs.** Training\n  in fp16 on hardware that doesn't have solid\n  BF16 support is a known source of loss spikes\n  and silent divergence. Force `bf16=True`\n  wherever the hardware supports it; don't fall\n  back to fp16 as if it were equivalent. Check\n  hardware support before picking a dtype:\n\n  ```bash\n  python -c \"import torch; print(torch.cuda.is_bf16_supported())\"\n  ```\n\n- **Rank too high on a small dataset overfits.**\n  A rank picked for \"SFT at scale\" (up to ~256)\n  on a dataset that doesn't have scale behind it\n  memorizes rather than generalizes. Match rank\n  to the Rank by Task table above, not to the\n  largest number available.\n- **Removing target modules to save memory costs\n  quality for negligible savings.** The adapter\n  parameters on `gate_proj`\u002F`up_proj`\u002F`down_proj`\n  are a small fraction of total model size — cutting\n  them barely moves memory but measurably hurts\n  quality. If memory is tight, move to QLoRA or\n  reduce rank\u002Fbatch\u002Fpack length before trimming\n  target modules.\n\nAll three failure modes share a pattern: they\nlook like a training-loop bug (loss spikes,\nplateaus, memorization) but are actually a\nconfig choice that contradicts the reference\nrecipe above. Check configuration against this\nskill before debugging the training loop itself.\n\n## References\n\n- `references\u002Fhyperparameters.md` — full rank\u002F\n  alpha\u002FLR tables by task type, rsLoRA notes,\n  batch\u002Fpacking interactions, and a complete\n  worked Unsloth config block.\n- `references\u002Funsloth-trl-mapping.md` — every\n  Unsloth kwarg mapped to its TRL\u002FPEFT\n  equivalent, current TRL API notes, and the\n  escape-hatch rule for when to drop back to\n  plain TRL.\n\nRelated skills: `finetuning-method-selection`\nroutes here; `dataset-curation` covers the data\nside this skill doesn't; `llm-finetuning-training-engineer`\nis the downstream consumer of the config this\nskill produces.\n",{"name":5,"description":7},"main","f846d9a97082ff30d613b55a085cef425d822473","plugins\u002Fllm-finetuning\u002Fskills\u002Flora-qlora-recipes","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle","https:\u002F\u002Fshotcowboystyle.github.io","Building smarter with AI, automation & modern tools. Exploring how far one person can go without a big team using today’s AI.","2026-11-10 04:58:27",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add shotcowboystyle\u002Fag-skills --skill lora-qlora-recipes",{"tool":38,"command":39},"bunx","bunx skills add shotcowboystyle\u002Fag-skills --skill lora-qlora-recipes",{"tool":41,"command":42},"pnpm","pnpm dlx skills add shotcowboystyle\u002Fag-skills --skill lora-qlora-recipes","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Flora-qlora-recipes\u002FSKILL.md",true,[46,52],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fhyperparameters.md","REFERENCE","参考文档",6801,"https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Flora-qlora-recipes\u002Freferences\u002Fhyperparameters.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Funsloth-trl-mapping.md",9884,"https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Fllm-finetuning\u002Fskills\u002Flora-qlora-recipes\u002Freferences\u002Funsloth-trl-mapping.md",[57,63,70,77,84,91],{"skillNo":58,"name":59,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":19},"01M7F1YRY5S5QRJ1V0BJS130QX","no-description","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fno-description\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M7F1Y9BV2FJNQJ3NB32JS5CM","good-skill","Fixture skill that satisfies every rule.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fgood-skill\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M7F1XXBQ3NQRXZSS0KAGAR9V","broken-link","Links to a reference file that does not exist.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fbroken-link\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7F1XBX4R46NR9W0HXQ7XCDX","Bad_Name","Name violates the kebab-case rule and does not match the directory.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fbad-name\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7F1WNAF8WNEQ7S5HKT5VK0D","skill-authoring","Write, review, and validate portable agent skills in SKILL.md format. Use when creating a new skill, editing an existing one, or asked whether a skill follows the Agent Skills spec.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fskills\u002Fskill-authoring\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":19},"01M7F1W30TC0XFNSF7YBRGETK2","weigh-bags","Record scale readings for each bag — empty (tare), outbound, current, final — and reconcile them against the itemised inventory to find unaccounted weight. Use when the user gives a bag weight, says how heavy something is, weighs their luggage, or asks how close to the limit they are. Appends to bags.yaml and runs scripts\u002Freconcile.py.","https:\u002F\u002Fgithub.com\u002Fshotcowboystyle\u002Fag-skills\u002Fblob\u002Fmain\u002Fplugins\u002Ftravel-packing-assistant\u002Fskills\u002Fweigh-bags\u002FSKILL.md",[],[]]