[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77WWFKD2N1B5D7K4ACS1AMN":3,"skill-same:01M77WWFKD2N1B5D7K4ACS1AMN":52},{"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":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M77WWFKD2N1B5D7K4ACS1AMN","trace-to-training-data",null,"Convert evaluation traces and production logs into SFT examples and preference pairs. Use when graded traces or failure examples exist and need to become training data, when applying rejection sampling to model outputs, or when building DPO pairs from passing and failing runs.","Reihantt6\u002Fsns-myagent","Reihantt6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F219822095?v=4","01M77KK94PV8B7C2D9XWXNS8DJ","Mohamad Reihan tegar prasetya","01M77KK958QFMTHYXGXFV7KZ79",2,"MIT","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Ftrace-to-training-data\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\ndescription: \"Convert evaluation traces and production logs into SFT examples and preference pairs. Use when graded traces or failure examples exist and need to become training data, when applying rejection sampling to model outputs, or when building DPO pairs from passing and failing runs.\"\n---\n# Trace To Training Data\n\nThis skill assumes `eval-harness-first`\nalready graded the traces being\nconverted here — goldens, graders,\nand `runs\u002F\u003Crun-id>\u002Fresults.json`\nall exist before conversion\nstarts. This is the flywheel edge\nthat skill names in its own flow:\n\"the same labeled traces become\nthe training set.\" Conversion\nhappens here; grading already\nhappened upstream.\n\n**Input:** graded traces —\n`eval\u002Fgoldens.jsonl` plus\n`runs\u002F\u003Crun-id>\u002Fresults.json`, each\nrow carrying a `task_id`, a\n`verdict` from the grader, and a\n`reward` when the task supports a\nscalar score (judge score,\nexecution partial-credit, or an\nRLVR verifier):\n\n```json\n{\"task_id\": \"t-042\", \"trace_id\": \"t-042-a3\",\n \"messages\": [{\"role\": \"user\", \"content\": \"...\"}],\n \"verdict\": \"pass\", \"reward\": 0.91,\n \"grader\": \"exact_match\"}\n```\n\n**Output format:** rows shaped\nexactly like `dataset-curation`'s\nFormat Selection table — SFT\n`messages` rows or DPO\n`prompt`\u002F`chosen`\u002F`rejected`\npairs — so this skill's output is\nthat skill's input with no\nreshaping step in between.\n\n## The Principle\n\nThe eval harness already did the\nlabeling work: every trace in\n`results.json` carries a verdict,\nand often a reward, before this\nskill ever touches it. Converting\na graded trace into a training\nrow is mechanical — pick a shape\nfrom `dataset-curation`'s table,\nmap fields, write JSONL.\n**Curation is the work that\nremains** — which traces clear a\nquality bar, which pairs are\ninformative, and which rows must\nnever enter the training set at\nall.\n\nTreat any conversion step that\nrequires re-judging a trace as a\nsign the harness is missing a\ngrader, not a gap this skill\nshould paper over. A trace with\nno verdict or reward isn't\nconvertible yet — route it back\nto `eval-harness-first` first,\ndon't hand-label it here to\nunblock conversion.\n\n## SFT From Traces\n\n- **Keep the top-reward fraction\n  of successful trajectories**,\n  not every passing one. Rank\n  passing traces by reward and\n  take a fraction (the\n  Agent-lightning pattern) rather\n  than every trace that merely\n  cleared the pass bar — a trace\n  that barely passed is a weaker\n  SFT signal than one that scored\n  well above threshold.\n- **Expert-corrected failures\n  become gold SFT examples\n  directly** (the Langfuse\n  pattern) — when a human edits a\n  failing trace's output into a\n  correct one, that correction\n  needs no reward threshold; a\n  human already validated it.\n  Route corrections straight into\n  the SFT set.\n- **Step-level masking beats\n  whole-trajectory discard for\n  multi-step traces.** When only\n  some steps in a multi-step\n  trajectory are bad, mask the\n  loss on the bad steps and keep\n  the good ones, rather than\n  discarding the whole trajectory.\n  SRFT reports 32.2% vs. 30.9% on\n  SWE-bench for step-level critic\n  masking over trajectory discard\n  — a real, if modest, gap from\n  the finer-grained cut.\n\n## Preference Pairs From Traces\n\n- **Build pairs from\n  passing-vs-failing trajectories\n  on the SAME task**, never from\n  unrelated best- and\n  worst-scoring traces pulled\n  across different tasks —\n  cross-task pairs teach the\n  model to prefer one task over\n  another, not one response over\n  another.\n- **Select the rejected member at\n  μ−2σ of the reward distribution\n  for that task, never the\n  absolute minimum.**\n  `preference-optimization`'s\n  Pair Construction section owns\n  the full selection formula;\n  this skill supplies the graded\n  trajectories it consumes.\n- **Judge-scored delta selection\n  cuts pair volume without\n  cutting signal.** Score each\n  candidate pair by\n  chosen-minus-rejected judge\n  delta and keep only the\n  highest-delta subset — the top\n  5k of a 16.5k candidate pool\n  matched the full pool's\n  downstream result. Build the\n  full candidate set first, then\n  filter by delta; don't cap\n  generation at 5k up front.\n\n## Hygiene\n\n- **Scan for secrets and PII before any row ships,\n  and redact what's found.** Traces sourced from\n  production logs can carry credentials, API keys,\n  tokens, or customer data — run a secret\u002FPII scan\n  over every SFT and DPO row and redact matches;\n  conversion fails closed (the row is dropped, not\n  shipped with the raw content) if sensitive fields\n  remain after redaction. Never commit secrets.\n- **Eval goldens must never leak\n  into training data.** Hold\n  every `eval\u002Fgoldens.jsonl` ID\n  out of every converted SFT and\n  DPO set — a trace that also\n  appears as a golden trains on\n  the exact item the checkpoint\n  gets graded against later,\n  silently inflating every\n  subsequent eval run.\n- **Dedup against the training\n  set**, not just within the\n  newly converted rows —\n  exact-match or\n  embedding-similarity, matching\n  `dataset-curation`'s dedup\n  method field, run against\n  whatever training data already\n  exists before this batch merges\n  in.\n- **Provenance goes into the\n  dataset card.** Every converted\n  row must trace back to its\n  source `run_id` and `trace_id`\n  — `dataset-curation`'s\n  Provenance field checks for\n  exactly this link back to\n  `trace-to-training-data`\n  output; a row with no traceable\n  source isn't ready to merge.\n\n## Related Skills\n\n- `eval-harness-first` — produces\n  the graded traces this skill\n  converts; a trace with no\n  verdict or reward isn't\n  convertible yet, route it back\n  there before conversion.\n- `dataset-curation` — owns the\n  target formats and the dataset\n  card this skill's provenance\n  data feeds; converted rows must\n  match its Format Selection\n  table field names exactly, not\n  an approximation of them.\n- `preference-optimization` —\n  consumes the DPO pairs this\n  skill builds and owns the full\n  μ−2σ rejection-selection\n  formula referenced above.\n\nWorked JSONL-to-JSONL conversions\n— graded trace to SFT row, trace\npair to DPO pair, correction to\nSFT row, the rejection-sampling\nloop, and the goldens-holdout\ncheck — live in\n`references\u002Fconversion-recipes.md`.\n","# Trace To Training Data\n\nThis skill assumes `eval-harness-first`\nalready graded the traces being\nconverted here — goldens, graders,\nand `runs\u002F\u003Crun-id>\u002Fresults.json`\nall exist before conversion\nstarts. This is the flywheel edge\nthat skill names in its own flow:\n\"the same labeled traces become\nthe training set.\" Conversion\nhappens here; grading already\nhappened upstream.\n\n**Input:** graded traces —\n`eval\u002Fgoldens.jsonl` plus\n`runs\u002F\u003Crun-id>\u002Fresults.json`, each\nrow carrying a `task_id`, a\n`verdict` from the grader, and a\n`reward` when the task supports a\nscalar score (judge score,\nexecution partial-credit, or an\nRLVR verifier):\n\n```json\n{\"task_id\": \"t-042\", \"trace_id\": \"t-042-a3\",\n \"messages\": [{\"role\": \"user\", \"content\": \"...\"}],\n \"verdict\": \"pass\", \"reward\": 0.91,\n \"grader\": \"exact_match\"}\n```\n\n**Output format:** rows shaped\nexactly like `dataset-curation`'s\nFormat Selection table — SFT\n`messages` rows or DPO\n`prompt`\u002F`chosen`\u002F`rejected`\npairs — so this skill's output is\nthat skill's input with no\nreshaping step in between.\n\n## The Principle\n\nThe eval harness already did the\nlabeling work: every trace in\n`results.json` carries a verdict,\nand often a reward, before this\nskill ever touches it. Converting\na graded trace into a training\nrow is mechanical — pick a shape\nfrom `dataset-curation`'s table,\nmap fields, write JSONL.\n**Curation is the work that\nremains** — which traces clear a\nquality bar, which pairs are\ninformative, and which rows must\nnever enter the training set at\nall.\n\nTreat any conversion step that\nrequires re-judging a trace as a\nsign the harness is missing a\ngrader, not a gap this skill\nshould paper over. A trace with\nno verdict or reward isn't\nconvertible yet — route it back\nto `eval-harness-first` first,\ndon't hand-label it here to\nunblock conversion.\n\n## SFT From Traces\n\n- **Keep the top-reward fraction\n  of successful trajectories**,\n  not every passing one. Rank\n  passing traces by reward and\n  take a fraction (the\n  Agent-lightning pattern) rather\n  than every trace that merely\n  cleared the pass bar — a trace\n  that barely passed is a weaker\n  SFT signal than one that scored\n  well above threshold.\n- **Expert-corrected failures\n  become gold SFT examples\n  directly** (the Langfuse\n  pattern) — when a human edits a\n  failing trace's output into a\n  correct one, that correction\n  needs no reward threshold; a\n  human already validated it.\n  Route corrections straight into\n  the SFT set.\n- **Step-level masking beats\n  whole-trajectory discard for\n  multi-step traces.** When only\n  some steps in a multi-step\n  trajectory are bad, mask the\n  loss on the bad steps and keep\n  the good ones, rather than\n  discarding the whole trajectory.\n  SRFT reports 32.2% vs. 30.9% on\n  SWE-bench for step-level critic\n  masking over trajectory discard\n  — a real, if modest, gap from\n  the finer-grained cut.\n\n## Preference Pairs From Traces\n\n- **Build pairs from\n  passing-vs-failing trajectories\n  on the SAME task**, never from\n  unrelated best- and\n  worst-scoring traces pulled\n  across different tasks —\n  cross-task pairs teach the\n  model to prefer one task over\n  another, not one response over\n  another.\n- **Select the rejected member at\n  μ−2σ of the reward distribution\n  for that task, never the\n  absolute minimum.**\n  `preference-optimization`'s\n  Pair Construction section owns\n  the full selection formula;\n  this skill supplies the graded\n  trajectories it consumes.\n- **Judge-scored delta selection\n  cuts pair volume without\n  cutting signal.** Score each\n  candidate pair by\n  chosen-minus-rejected judge\n  delta and keep only the\n  highest-delta subset — the top\n  5k of a 16.5k candidate pool\n  matched the full pool's\n  downstream result. Build the\n  full candidate set first, then\n  filter by delta; don't cap\n  generation at 5k up front.\n\n## Hygiene\n\n- **Scan for secrets and PII before any row ships,\n  and redact what's found.** Traces sourced from\n  production logs can carry credentials, API keys,\n  tokens, or customer data — run a secret\u002FPII scan\n  over every SFT and DPO row and redact matches;\n  conversion fails closed (the row is dropped, not\n  shipped with the raw content) if sensitive fields\n  remain after redaction. Never commit secrets.\n- **Eval goldens must never leak\n  into training data.** Hold\n  every `eval\u002Fgoldens.jsonl` ID\n  out of every converted SFT and\n  DPO set — a trace that also\n  appears as a golden trains on\n  the exact item the checkpoint\n  gets graded against later,\n  silently inflating every\n  subsequent eval run.\n- **Dedup against the training\n  set**, not just within the\n  newly converted rows —\n  exact-match or\n  embedding-similarity, matching\n  `dataset-curation`'s dedup\n  method field, run against\n  whatever training data already\n  exists before this batch merges\n  in.\n- **Provenance goes into the\n  dataset card.** Every converted\n  row must trace back to its\n  source `run_id` and `trace_id`\n  — `dataset-curation`'s\n  Provenance field checks for\n  exactly this link back to\n  `trace-to-training-data`\n  output; a row with no traceable\n  source isn't ready to merge.\n\n## Related Skills\n\n- `eval-harness-first` — produces\n  the graded traces this skill\n  converts; a trace with no\n  verdict or reward isn't\n  convertible yet, route it back\n  there before conversion.\n- `dataset-curation` — owns the\n  target formats and the dataset\n  card this skill's provenance\n  data feeds; converted rows must\n  match its Format Selection\n  table field names exactly, not\n  an approximation of them.\n- `preference-optimization` —\n  consumes the DPO pairs this\n  skill builds and owns the full\n  μ−2σ rejection-selection\n  formula referenced above.\n\nWorked JSONL-to-JSONL conversions\n— graded trace to SFT row, trace\npair to DPO pair, correction to\nSFT row, the rejection-sampling\nloop, and the goldens-holdout\ncheck — live in\n`references\u002Fconversion-recipes.md`.\n",{"description":7},"main","58843b159d7a4050d6b785a98b908cce505c0b92","agent\u002Fskills\u002Ftrace-to-training-data",0,"https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent","https:\u002F\u002Fgithub.com\u002FReihantt6","reihsns.web.id","2026-08-19 20:08:18",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add Reihantt6\u002Fsns-myagent --skill trace-to-training-data",{"tool":38,"command":39},"bunx","bunx skills add Reihantt6\u002Fsns-myagent --skill trace-to-training-data",{"tool":41,"command":42},"pnpm","pnpm dlx skills add Reihantt6\u002Fsns-myagent --skill trace-to-training-data","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Ftrace-to-training-data\u002FSKILL.md",true,[46],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"references\u002Fconversion-recipes.md","REFERENCE","参考文档",7770,"https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Ftrace-to-training-data\u002Freferences\u002Fconversion-recipes.md",[53,60,67,74,81,88],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M77XDCYZ407XYMF2T4P6F7VT","youtube-transcript","Extract transcripts from YouTube videos. Use when the user asks for a transcript, subtitles, or captions of a YouTube video and provides a YouTube URL (youtube.com\u002Fwatch?v=, youtu.be\u002F, or similar). Supports output with or without timestamps.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-transcript\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M77XD7PBC98J6MXGJJ2TJ8DX","youtube-thumbnail-design","YouTube thumbnail design with specific dimensions, contrast rules, and mobile preview optimization. Covers safe zones, text placement, face expression psychology, and A\u002FB testing. Use for: YouTube thumbnails, video cover images, click-through optimization. Triggers: youtube thumbnail, thumbnail design, video thumbnail, click through rate, ctr optimization, youtube cover, video cover image, thumbnail maker, thumbnail tips, youtube design, video preview image","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-thumbnail-design\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M77XD2541C8HXJ6RHYZP4NGJ","youtube-short-maker","Make YouTube Shorts that hold attention with Pexo — vertical 9:16, fast-paced, hook in the first 3 seconds. Describe your topic and Pexo writes the script, generates the shots, picks the models, and assembles a finished Short with music and captions. Use for YouTube Shorts: \"youtube short\", \"make a youtube short\", \"shorts video\", \"short-form video\". NOT for long-form or landscape video.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fyoutube-short-maker\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M77XCXKVSYK7KTZ3PQBVVGFD","xcode-project-setup","Safely modifies Xcode projects (.pbxproj) to add Swift Packages and link files. Use this skill whenever an iOS project needs dependencies installed (e.g. Firebase, Alamofire).","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fxcode-project-setup\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M77XCR1DV8ZQK2NHWKVY4EWX","x402","Set up Browser Use Cloud payments with x402 — pay per request from a crypto wallet (USDC on Base mainnet), no signup or API key. Two setups it works out up front — \"just use it\" (set up a wallet so you or Claude Code can run cloud browser tasks paid from the wallet — Claude writes and runs throwaway scripts, nothing touches your codebase) or \"build it in\" (install the SDK and write the key + code into your project). Walks through wallet setup, funding, .env, and a ~$1 test run. Use when the user asks about x402, pay-per-use, USDC payments, or wants Browser Use Cloud without an API key. For the free-tier signup (reverse-CAPTCHA → API key), use `browser-use cloud signup` or the `cloud` skill instead.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fx402\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M77XCMD3R3ZBZQQSW6E0793B","x-twitter-scraper","Build GitHub Copilot workflows with Xquik X API SDKs, REST endpoints, hosted Apify Actor runs, MCP tools, TweetClaw OpenClaw plugin installs, signed webhooks, tweet search, user lookup, follower exports, media actions, and agent automation.","https:\u002F\u002Fgithub.com\u002FReihantt6\u002Fsns-myagent\u002Fblob\u002Fmain\u002Fagent\u002Fskills\u002Fx-twitter-scraper\u002FSKILL.md",[],[]]