[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M796J45G7E1V9PM0DFCPDWWC":3,"skill-same:01M796J45G7E1V9PM0DFCPDWWC":74},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":6,"ownerHtmlUrl":31,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":32,"install":33,"resources":45},"01M796J45G7E1V9PM0DFCPDWWC","train-llm-from-scratch",null,"Route train-llm-from-scratch workflows for from-scratch PyTorch LLM pretraining, SFT, reward modeling, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR, evaluation, chat, configs, UI, and data validation.","VectorSpaceLab\u002FAREX-Skill","VectorSpaceLab","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F180637102?v=4","01M7538C5WPSJGWF7KHR4EFZ9F","01M7538C6R2GFP53NPXTCZNBRT",351,"MIT","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: train-llm-from-scratch\ndescription: \"Route train-llm-from-scratch workflows for from-scratch PyTorch\n  LLM pretraining, SFT, reward modeling, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR,\n  evaluation, chat, configs, UI, and data validation.\"\ndisable-model-invocation: true\nmetadata:\n  disco-role: operating\nlicense: MIT\n---\n\n# train-llm-from-scratch\n\nUse this repo skill when a task involves the `train-llm-from-scratch` project: a\npure-PyTorch educational LLM stack covering raw text tokenization, decoder-only\nTransformer pretraining, SFT, reward modeling, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR,\nGSM8K evaluation, chat, JSON configs, and a Streamlit control panel.\n\n## Start Here\n\n1. If the task may depend on repo freshness, read\n   [`references\u002Frepo-provenance.md`](references\u002Frepo-provenance.md).\n2. For setup, optional extras, backend expectations, and import checks, read\n   [`references\u002Finstallation.md`](references\u002Finstallation.md).\n3. Run [`scripts\u002Fcheck_environment.py`](scripts\u002Fcheck_environment.py) when you\n   need a safe import\u002Fbackend smoke check in the user's environment.\n4. If checkpoint contents or model dimensions are unclear, run\n   [`scripts\u002Finspect_checkpoint.py`](scripts\u002Finspect_checkpoint.py) before\n   loading the model.\n5. Route to the narrowest sub-skill below. Keep this root as the router; do not\n   bury detailed stage work here.\n\n## Sub-Skill Routing\n\n| User task | Read |\n|---|---|\n| Prepare or validate Pile HDF5, packed SFT HDF5, preference JSONL, RL prompt JSONL, arithmetic curriculum, tokenizer\u002Fmask\u002Fdata-loader issues | [`sub-skills\u002Fdata-preparation\u002FSKILL.md`](sub-skills\u002Fdata-preparation\u002FSKILL.md) |\n| Inspect the custom Transformer, plan base pretraining, choose legacy vs modern pretraining, DDP\u002Fbf16, memory, checkpoint\u002Fresume, tiny model smoke | [`sub-skills\u002Fmodel-pretraining\u002FSKILL.md`](sub-skills\u002Fmodel-pretraining\u002FSKILL.md) |\n| Run or debug SFT, reward model, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR, rollout\u002Flog-probs, KL\u002Freward metrics, post-training pipeline sequencing | [`sub-skills\u002Fpost-training-rlhf\u002FSKILL.md`](sub-skills\u002Fpost-training-rlhf\u002FSKILL.md) |\n| Evaluate on GSM8K, build stage tables, chat with checkpoints, choose raw vs chat mode, sampling controls, answer parsing\u002Fverifier behavior | [`sub-skills\u002Fevaluation-chat\u002FSKILL.md`](sub-skills\u002Fevaluation-chat\u002FSKILL.md) |\n| Edit JSON configs, use smoke configs, inspect CLI override precedence, operate the Streamlit UI, inspect metrics JSONL\u002Fjob logs | [`sub-skills\u002Fconfiguration-ui\u002FSKILL.md`](sub-skills\u002Fconfiguration-ui\u002FSKILL.md) |\n\n## Install And Import Baseline\n\nThe package metadata exposes import roots `config`, `data_loader`, `src`, and\n`ui`. A typical editable install from a compatible checkout is:\n\n```bash\npip install -e .\npip install -e \".[train]\"   # datasets + optional wandb for data\u002Fpost-training\npip install -e \".[ui]\"      # streamlit + pandas + altair for the control panel\npip install -e \".[docs]\"    # mkdocs tooling, only for docs-site work\n```\n\nFor the selected training workflows, use a PyTorch build that matches the\nmachine. Full training is CUDA-oriented; CPU smoke configs are useful for parser,\nconfig, and algorithm checks but do not verify large CUDA\u002Fbf16\u002FDDP training.\n\nMinimal import smoke:\n\n```bash\npython - \u003C\u003C'PY'\nfrom src.models.transformer import Transformer\nfrom src.post_training.chat_template import encode_chat, EOT_ID\nids, mask = encode_chat([\n    {\"role\": \"user\", \"content\": \"What is 2+2?\"},\n    {\"role\": \"assistant\", \"content\": \"\u003Canswer>4\u003C\u002Fanswer>\"},\n])\nprint(len(ids), len(mask), sum(mask), ids[-1] == EOT_ID)\nprint(Transformer(n_head=2, n_embed=16, context_length=8, vocab_size=64, N_BLOCKS=1).__class__.__name__)\nPY\n```\n\n## Common Operating Rules\n\n- Start with smoke configs or dry-run command builders before launching long\n  training, data downloads, or multi-GPU jobs.\n- Validate data files before training. Mask\u002Fdata errors silently corrupt SFT,\n  preference optimization, and RL rewards.\n- Treat `\u002Fephemeral`-style paths in examples as user-configurable storage roots,\n  not mandatory constants.\n- Keep Weights & Biases optional. The repo writes local JSONL metrics even when\n  W&B is disabled or unavailable.\n- Use greedy decoding for comparable GSM8K stage tables; sampling is for chat.\n- Do not claim CUDA coverage from CPU-only smoke tests. For CUDA claims, verify a\n  torch CUDA allocation and at least a tiny model forward on the target host.\n\n## Troubleshooting\n\nRead [`references\u002Ftroubleshooting.md`](references\u002Ftroubleshooting.md) for\ncross-cutting install\u002Fimport\u002Fbackend\u002Fcheckpoint routing. Workflow-specific\nfailures live in each sub-skill's nearest `references\u002Ftroubleshooting.md`.\n","# train-llm-from-scratch\n\nUse this repo skill when a task involves the `train-llm-from-scratch` project: a\npure-PyTorch educational LLM stack covering raw text tokenization, decoder-only\nTransformer pretraining, SFT, reward modeling, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR,\nGSM8K evaluation, chat, JSON configs, and a Streamlit control panel.\n\n## Start Here\n\n1. If the task may depend on repo freshness, read\n   [`references\u002Frepo-provenance.md`](references\u002Frepo-provenance.md).\n2. For setup, optional extras, backend expectations, and import checks, read\n   [`references\u002Finstallation.md`](references\u002Finstallation.md).\n3. Run [`scripts\u002Fcheck_environment.py`](scripts\u002Fcheck_environment.py) when you\n   need a safe import\u002Fbackend smoke check in the user's environment.\n4. If checkpoint contents or model dimensions are unclear, run\n   [`scripts\u002Finspect_checkpoint.py`](scripts\u002Finspect_checkpoint.py) before\n   loading the model.\n5. Route to the narrowest sub-skill below. Keep this root as the router; do not\n   bury detailed stage work here.\n\n## Sub-Skill Routing\n\n| User task | Read |\n|---|---|\n| Prepare or validate Pile HDF5, packed SFT HDF5, preference JSONL, RL prompt JSONL, arithmetic curriculum, tokenizer\u002Fmask\u002Fdata-loader issues | [`sub-skills\u002Fdata-preparation\u002FSKILL.md`](sub-skills\u002Fdata-preparation\u002FSKILL.md) |\n| Inspect the custom Transformer, plan base pretraining, choose legacy vs modern pretraining, DDP\u002Fbf16, memory, checkpoint\u002Fresume, tiny model smoke | [`sub-skills\u002Fmodel-pretraining\u002FSKILL.md`](sub-skills\u002Fmodel-pretraining\u002FSKILL.md) |\n| Run or debug SFT, reward model, DPO\u002FORPO\u002FKTO, PPO, GRPO\u002FRLVR, rollout\u002Flog-probs, KL\u002Freward metrics, post-training pipeline sequencing | [`sub-skills\u002Fpost-training-rlhf\u002FSKILL.md`](sub-skills\u002Fpost-training-rlhf\u002FSKILL.md) |\n| Evaluate on GSM8K, build stage tables, chat with checkpoints, choose raw vs chat mode, sampling controls, answer parsing\u002Fverifier behavior | [`sub-skills\u002Fevaluation-chat\u002FSKILL.md`](sub-skills\u002Fevaluation-chat\u002FSKILL.md) |\n| Edit JSON configs, use smoke configs, inspect CLI override precedence, operate the Streamlit UI, inspect metrics JSONL\u002Fjob logs | [`sub-skills\u002Fconfiguration-ui\u002FSKILL.md`](sub-skills\u002Fconfiguration-ui\u002FSKILL.md) |\n\n## Install And Import Baseline\n\nThe package metadata exposes import roots `config`, `data_loader`, `src`, and\n`ui`. A typical editable install from a compatible checkout is:\n\n```bash\npip install -e .\npip install -e \".[train]\"   # datasets + optional wandb for data\u002Fpost-training\npip install -e \".[ui]\"      # streamlit + pandas + altair for the control panel\npip install -e \".[docs]\"    # mkdocs tooling, only for docs-site work\n```\n\nFor the selected training workflows, use a PyTorch build that matches the\nmachine. Full training is CUDA-oriented; CPU smoke configs are useful for parser,\nconfig, and algorithm checks but do not verify large CUDA\u002Fbf16\u002FDDP training.\n\nMinimal import smoke:\n\n```bash\npython - \u003C\u003C'PY'\nfrom src.models.transformer import Transformer\nfrom src.post_training.chat_template import encode_chat, EOT_ID\nids, mask = encode_chat([\n    {\"role\": \"user\", \"content\": \"What is 2+2?\"},\n    {\"role\": \"assistant\", \"content\": \"\u003Canswer>4\u003C\u002Fanswer>\"},\n])\nprint(len(ids), len(mask), sum(mask), ids[-1] == EOT_ID)\nprint(Transformer(n_head=2, n_embed=16, context_length=8, vocab_size=64, N_BLOCKS=1).__class__.__name__)\nPY\n```\n\n## Common Operating Rules\n\n- Start with smoke configs or dry-run command builders before launching long\n  training, data downloads, or multi-GPU jobs.\n- Validate data files before training. Mask\u002Fdata errors silently corrupt SFT,\n  preference optimization, and RL rewards.\n- Treat `\u002Fephemeral`-style paths in examples as user-configurable storage roots,\n  not mandatory constants.\n- Keep Weights & Biases optional. The repo writes local JSONL metrics even when\n  W&B is disabled or unavailable.\n- Use greedy decoding for comparable GSM8K stage tables; sampling is for chat.\n- Do not claim CUDA coverage from CPU-only smoke tests. For CUDA claims, verify a\n  torch CUDA allocation and at least a tiny model forward on the target host.\n\n## Troubleshooting\n\nRead [`references\u002Ftroubleshooting.md`](references\u002Ftroubleshooting.md) for\ncross-cutting install\u002Fimport\u002Fbackend\u002Fcheckpoint routing. Workflow-specific\nfailures live in each sub-skill's nearest `references\u002Ftroubleshooting.md`.\n",{"name":5,"license":14,"metadata":23,"description":7,"disable-model-invocation":25},{"disco-role":24},"operating",true,"main","ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6","skills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch",27,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab","2026-09-03 10:25:21",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":25},[35,38,41],{"tool":36,"command":37},"npx","npx skills add VectorSpaceLab\u002FAREX-Skill --skill train-llm-from-scratch",{"tool":39,"command":40},"bunx","bunx skills add VectorSpaceLab\u002FAREX-Skill --skill train-llm-from-scratch",{"tool":42,"command":43},"pnpm","pnpm dlx skills add VectorSpaceLab\u002FAREX-Skill --skill train-llm-from-scratch","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002FSKILL.md",[46,52,56,60,64,70],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Finstallation.md","REFERENCE","参考文档",4260,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Freferences\u002Finstallation.md",{"path":53,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":54,"sourceHtmlUrl":55},"references\u002Frepo-provenance.md",2750,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Freferences\u002Frepo-provenance.md",{"path":57,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":58,"sourceHtmlUrl":59},"references\u002Frepo-routing-metadata.json",598,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Freferences\u002Frepo-routing-metadata.json",{"path":61,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":62,"sourceHtmlUrl":63},"references\u002Ftroubleshooting.md",4666,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Freferences\u002Ftroubleshooting.md",{"path":65,"resourceType":66,"resourceTypeDesc":67,"sizeBytes":68,"sourceHtmlUrl":69},"scripts\u002Fcheck_environment.py","SCRIPT","脚本",5724,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Fscripts\u002Fcheck_environment.py",{"path":71,"resourceType":66,"resourceTypeDesc":67,"sizeBytes":72,"sourceHtmlUrl":73},"scripts\u002Finspect_checkpoint.py",4625,"https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Frepositories\u002Frepo-skills\u002Ftrain-llm-from-scratch\u002Fscripts\u002Finspect_checkpoint.py",[75,83,90,97,104,111],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":80,"categories":81,"keywords":82,"publishedAt":18},"01M799NYSH1R3RVAHSM2D88Y84","passnet-triton-opt","Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block\u002Fgrid\u002Fwarp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.","Apache-2.0","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-triton-opt\u002FSKILL.md",[],[],{"skillNo":84,"name":85,"nameZh":6,"description":86,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":87,"categories":88,"keywords":89,"publishedAt":18},"01M799NS5APEC0N1FZ7XZVW6BN","passnet-solve","END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-solve\u002FSKILL.md",[],[],{"skillNo":91,"name":92,"nameZh":6,"description":93,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":94,"categories":95,"keywords":96,"publishedAt":18},"01M799NNY9BKSM21V5R7QVTBJX","passnet-skill","PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under .\u002Fpass_dir\u002F, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-skill\u002FSKILL.md",[],[],{"skillNo":98,"name":99,"nameZh":6,"description":100,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":101,"categories":102,"keywords":103,"publishedAt":18},"01M799NHH6NP7Z1RK77KGXAN6S","passnet-pattern-fusion","Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()\u002Freplacement_args()\u002Freplacement_func() files, when a pass \"failed to match\", or when deciding how to fuse multiple kernels\u002Fops.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-pattern-fusion\u002FSKILL.md",[],[],{"skillNo":105,"name":106,"nameZh":6,"description":107,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":108,"categories":109,"keywords":110,"publishedAt":18},"01M799ND2J4WNC2PGAZ0A06NDF","passnet-orchestrate","Round\u002Fbatch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep\u002Frevert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-orchestrate\u002FSKILL.md",[],[],{"skillNo":112,"name":113,"nameZh":6,"description":114,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":79,"licenseFileUrl":6,"sourceHtmlUrl":115,"categories":116,"keywords":117,"publishedAt":18},"01M799N8SBTMY02ZDG2AZF5ZWE","passnet-feedback","Fast, deterministic feedback for PassNet work: pre-flight pattern\u002Fmatch verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).","https:\u002F\u002Fgithub.com\u002FVectorSpaceLab\u002FAREX-Skill\u002Fblob\u002Fmain\u002Fskills\u002Ftask-oriented\u002FPassNet\u002Fskills\u002Fpassnet-feedback\u002FSKILL.md",[],[]]