[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7CQCPDK0SPCFZEBCQFZ362C":3,"skill-same:01M7CQCPDK0SPCFZEBCQFZ362C":58},{"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":36,"commitSha":37,"skillDirectory":38,"forksCount":39,"repositoryHtmlUrl":40,"repositoryHomepageUrl":6,"ownerHtmlUrl":41,"ownerBlogUrl":6,"ownerDescription":42,"updatedAt":43,"install":44,"resources":57},"01M7CQCPDK0SPCFZEBCQFZ362C","tao-finetune-huggingface-model",null,"Fine-tune any HuggingFace CV \u002F VLM \u002F LLM model on local NVIDIA GPUs inside an NGC PyTorch container. Use when the user wants to fine-tune a HuggingFace model (full or LoRA), train a vision \u002F VLM \u002F LLM model end-to-end, generate a reproducible HF training pipeline, smoke-test a HuggingFace model locally before scale-up, push a fine-tuned model to the HF Hub with a model card, or emit a self-contained rerun skill for an existing HuggingFace finetune. Supports image classification, object detection, semantic \u002F instance \u002F panoptic segmentation, depth estimation, image-text-to-text VLM (SFT \u002F LoRA), and LLM SFT \u002F DPO \u002F GRPO. Six-step workflow: inspect and qualify, hardware and NGC image, research, generate and smoke, train + eval + infer, push and emit rerun skill.","om-scogo\u002Fskillsh-scraper","om-scogo","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F160466840?v=4","01M79AWB6DFVCQ7RK70CS34B96","Om Achrekar","01M79AWB6ZM9SKR8KDH1WSSM22",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fnvidia-skills\u002Ftao-finetune-huggingface-model\u002Fskills\u002Ftao-finetune-huggingface-model\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: tao-finetune-huggingface-model\ndescription: >\n  Fine-tune any HuggingFace CV \u002F VLM \u002F LLM model on local NVIDIA GPUs inside an\n  NGC PyTorch container. Use when the user wants to fine-tune a HuggingFace\n  model (full or LoRA), train a vision \u002F VLM \u002F LLM model end-to-end, generate a\n  reproducible HF training pipeline, smoke-test a HuggingFace model locally\n  before scale-up, push a fine-tuned model to the HF Hub with a model card, or\n  emit a self-contained rerun skill for an existing HuggingFace finetune.\n  Supports image classification, object detection, semantic \u002F instance \u002F\n  panoptic segmentation, depth estimation, image-text-to-text VLM (SFT \u002F LoRA),\n  and LLM SFT \u002F DPO \u002F GRPO. Six-step workflow: inspect and qualify, hardware\n  and NGC image, research, generate and smoke, train + eval + infer, push and\n  emit rerun skill.\nlicense: Apache-2.0\ntags:\n  - finetuning\n  - huggingface\n  - nvidia-tao\n  - computer-vision\n  - training\ncompatibility: Requires docker + nvidia-container-toolkit, NVIDIA GPU (driver ≥ 545, ≥ 24 GB VRAM for ≤3B models), ~40 GB free disk. Optional credentials (read from the session environment, exported before launching) — HF_TOKEN is read only when the model\u002Fdataset is gated or `push_to_hub` is on; WANDB_API_KEY and WANDB_PROJECT only when WandB logging is enabled.\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\nallowed-tools: Read Bash Write\n---\n\u003C!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0; see http:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0 -->\n\n# tao-finetune-huggingface-model\n\nLocal NVIDIA GPU fine-tuning for HuggingFace models, grounded in live-fetched\ndocumentation with curated references as a fallback safety net. One NGC container,\na few focused scripts, one push to HF Hub. Follow the rules in this file; don't\nimprovise.\n\n**Order of authority (highest first):**\n\n1. **User input** — explicit `model_id`, `dataset_id`, `training_method`, `config.yaml` overrides.\n2. **Live research** — model card, HF repo example, author finetune script, HF task docs, paper; always fetched (Step 3 + `references\u002Fresearch-priorities.md`).\n3. **Curated references** (`references\u002F*.md`) — fallback when live research is silent\u002Fambiguous.\n4. **Your training-data memory** — last resort; suspect, cross-check against (2)\u002F(3).\n\nConflict resolution between (2) and (3) and the source-line discrepancy note are\nin `references\u002Fresearch-priorities.md`.\n\n---\n\n## Inputs\n\n**Required:**\n- `model_id` — HuggingFace model ID, e.g. `google\u002Fvit-base-patch16-224`\n\n**Conditional credentials (read from the session environment, exported before launching when present):**\n- `HF_TOKEN` — only when the model\u002Fdataset is **gated** (read) or `push_to_hub` is on (write); public + public + `push_to_hub: false` needs none. Value never read — presence-only via `[ -n \"$HF_TOKEN\" ]`.\n- `WANDB_API_KEY`, `WANDB_PROJECT` — only when WandB is enabled; `WANDB_MODE=disabled` opts out.\n\n**Dataset — exactly one:**\n- `dataset_id` — HuggingFace dataset ID *(source: `hf`)*\n- `local_dataset_path` — local folder or file *(source: `local`)*; optional\n  `local_dataset_format` ∈ {auto, imagefolder, coco, voc, jsonl, arrow, parquet,\n  csv} (default: auto-detect).\n- *(omit)* — agent recommends popular datasets *(source: `recommend`)*\n\n**Optional (have defaults):**\n- `task_type` — auto-detected from config + model card\n- `n_train=10000`, `n_eval=1000`, `n_epochs=3`, `lora_r=16`\n- `output_dir=.\u002Foutput\u002F\u003Cmodel_short_name>`\n- `hf_model_repo` — push target; if unset and HF_TOKEN has write access,\n  auto-derived as `\u003Cwhoami>\u002F\u003Cmodel_short_name>-finetuned`.\n- `push_to_hub=True` — set to `False` to skip\n- `skip_baseline=False` — skip zero-shot baseline eval\n\n**Optional deliverables (off by default):**\n```yaml\nemit_progress_log: false   # output_dir\u002FPROGRESS.md (per-step journal)\nemit_report:       false   # reports\u002Freport.{pdf,html} with curves & samples\nemit_unit_tests:   false   # tests\u002F with fake-data heterogeneous-batch tests\n```\n\nAll values live in `output_dir\u002Fconfig.yaml`. Never hardcode in Python.\n\n---\n\n## Execution platform\n\nThis skill orchestrates *what* to run; the platform skills own *how* to run it on\na GPU host — read them first.\n\n| Concern | Authoritative skill |\n|---|---|\n| GPU host runtime (driver 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit 1.19.0) | [`tao-skill-bank:tao-setup-nvidia-gpu-host`](..\u002F..\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002FSKILL.md) |\n| `docker run` flags, NGC auth, mounts, env passthrough | [`tao-skill-bank:tao-run-on-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-docker\u002FSKILL.md) |\n| Local Docker job preflight (daemon, GPU smoke) | [`tao-skill-bank:tao-run-on-local-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-local-docker\u002FSKILL.md) |\n\n**Default platform:** `local-docker` — build a one-off image (`run-\u003Cshort>:latest`)\nand run it on the local Docker daemon. Ask only when the user explicitly needs a\ndifferent backend (Brev remote GPU, SLURM\u002FKubernetes); then run that platform's\nPreflight first and route the Steps 4–5 `docker run` commands through it. The\nGPU-runtime and presence-only credential preflights (values never read), the\ncanonical `docker run` flag set, the `list_tao_platforms.py` selection command, and\nthe workflow-specific flags (`--entrypoint \u002Fbin\u002Fbash -lc`, `PYTORCH_CUDA_ALLOC_CONF`,\n`--name hft_train`) are in `references\u002Fworkflow-intake-preflight.md`.\n\n---\n\n## References — fallback safety net\n\nConsulted **only** when live research is silent, ambiguous, or unavailable; live\ndocs always win for the specific model and current API. Each step links the\nreferences it needs; full catalog in `references\u002Fdetailed-workflow.md`.\n\nAlways-on: `core-rules.md`, `error-playbook.md`, `compat-workarounds.md`,\n`model-discovery.md`, `dataset-recommendations.md`, `dataset-sources.md`,\n`dataset-patterns.md`, `hardware-container.md`, `research-priorities.md`,\n`cv-scripts.md`, `vlm-scripts.md`, `docker-runs.md`, `hub-push.md`,\n`pipeline-skill-template.md`, `deliverables.md`. Opt-in (when their flag\u002Fneed\napplies): `progress-tracking.md`, `testing.md`, `reporting.md`,\n`workflow-intake-preflight.md`, `workflow-generate-train.md`, `workflow-push-rerun.md`.\n\n**Rule:** before falling back, log the live source you tried and why it was\ninsufficient (`config.yaml` `notes:`, and PROGRESS.md if enabled). `[FETCH LIVE]`\nmarkers in `cv-scripts.md` \u002F `vlm-scripts.md` are a research checklist, not code to\ninline — refetch the listed URL if a block has no Step 3 finding.\n\n---\n\n## Core rules\n\nNon-negotiable behaviors. **Short version** (full enumeration —\nhallucinated-imports list, never-without-approval list, full error-recovery and\nhardware-sizing tables — in `references\u002Fcore-rules.md`, consult before any\ntraining-time decision):\n\n- **Your HF-library knowledge is outdated.** Fetch live docs (model card, HF\n  repo example, task doc) before writing any ML code — don't generate trainer\n  args \u002F collator \u002F transforms from memory (Step 3).\n- **Smoke-test on real data with `--max_steps 1`** before any full run; no batch\n  launches without a verified smoke.\n- **Never silently substitute** model_id, dataset_id, or training_method — if\n  what the user asked for doesn't load, stop and ask.\n- **Error recovery is minimal-change.** OOM → halve batch, double grad_accum,\n  enable gradient checkpointing (no LoRA switch without approval); NaN → reduce\n  LR 10×; flat loss → inspect collator; same error 3× → stop and ask. Don't loop.\n- **Dataset columns verified BEFORE the collator** — rename in `prepare_data.py`;\n  restructuring needed → stop and ask.\n- **Hardware-sizing thumb (bf16):** ≤3B → 24 GB, 7–13B → 80 GB, 30B+ → multi-GPU\n  or LoRA on 1× 80 GB, 70B+ → 8× 80 GB or LoRA. Full finetune won't fit and no\n  LoRA requested → ask before switching.\n\n---\n\n## Workflow — 6 steps\n\nSingle pass, sequential; each step has a clear gate before the next begins.\n\n### Step 1 — Inspect & qualify\n\n**Goal:** decide whether to proceed. Probe model + dataset, apply accept\u002Freject,\nregister applicable compat fixes, write the initial `config.yaml`.\n\nPrerequisites: `MODEL_ID`, optional `DATASET_ID` \u002F `local_dataset_path`,\noptional `HF_TOKEN`, `OUTPUT_DIR` (default `.\u002Foutput\u002F\u003Cmodel_short_name>`). Probes\nrun in a CPU-only `python:3.12-slim` Docker container (bind-mounted `.probe\u002F`\nscratch) so the host needs no virtualenv — Docker must exist first. Docker-presence\nguard, container env, full probe invocation, and the model\u002Fdataset probe scripts\nare in `references\u002Fworkflow-intake-preflight.md`, `references\u002Fmodel-discovery.md`,\nand `references\u002Fdataset-sources.md`.\n\nProbe requirements:\n\n- Model: load `AutoConfig`, read model-card tags, detect task from\n  `architectures` + tags + card examples (fallback logging in `model-discovery.md`).\n- Dataset: for recommended datasets, first present 3-5 choices from\n  `dataset-recommendations.md`; for local data, bind-mount read-only and use\n  `dataset-sources.md` format detection.\n- Reject early if the model config fails, the task is out of scope, no recipe\n  source exists, or the dataset cannot load \u002F match the task schema.\n- Evaluate `compat-workarounds.md` against the model\u002Ftask; defer hardware-dependent\n  rules to Step 2.\n\nWrite the initial `config.yaml` (`model_id`, `task`, `dataset_id` or\n`local_dataset_path`, `research_sources: []` filled in Step 3,\n`applicable_workarounds:` from Step 1, `notes: []` for reference fallbacks,\n`push_to_hub: true` default — annotated template in\n`references\u002Fworkflow-intake-preflight.md`). Optionally `rm -rf \"$OUTPUT_DIR\u002F.probe\"`\nonce the gate is met.\n\n**Gate:** `config.yaml` exists with model, dataset, task, applicable_workarounds;\ndo not proceed if any field is missing.\n\n---\n\n### Step 2 — Hardware audit & NGC image\n\n**Goal:** verify Docker + GPU + disk, pick the NGC PyTorch image live, finalize\nhardware-dependent compat rules.\n\n**2a. Audit (hard gate)** — three checks (commands in\n`references\u002Fworkflow-intake-preflight.md`):\n1. GPU host runtime — `tao-setup-nvidia-gpu-host`'s\n   `setup-nvidia-gpu-host.sh --backend docker --check-only`; on fail, ask approval\n   then re-run with `--install --yes`.\n2. Free-disk soft-warn — override via `MIN_DISK_GB` (default 100 GB); recommend\n   ≥ 100 GB for NGC base (~20 GB) + HF cache + checkpoints + data.\n3. Conditional credential presence (from the session environment, values never\n   read) — `HF_TOKEN` only when gated or `push_to_hub` is on; `WANDB_*` only when\n   WandB is on.\n\n**Do not proceed to Step 4 on a hard-fail** — Step 4's `docker build` pulls a\n20+ GB NGC base, and a missing `nvidia-container-toolkit` only surfaces later as\n`could not select device driver \"\" with capabilities: [[gpu]]`. Record `gpu_count`,\n`gpu_name`, `driver_major`, `vram_gb_per_gpu` in `config.yaml`.\n\n**2b. Pick NGC image (live):** from the NVIDIA Deep Learning Frameworks support\nmatrix (\u003Chttps:\u002F\u002Fdocs.nvidia.com\u002Fdeeplearning\u002Fframeworks\u002Fsupport-matrix\u002Findex.html>),\nPyTorch NGC container section, pick the highest-versioned image where\n`Min driver ≤ detected driver_major` and container CUDA `≤` host CUDA Toolkit\n(match closely so cuDNN \u002F TensorRT line up). Do **not** reject an image for an\n`aN`\u002F`bN`\u002F`rcN` PyTorch tag — NGC validates the full image; pick the newest\nCUDA-aligned one and let `compat-workarounds.md` handle per-version issues. If the\nmatrix is unreachable, use the fallbacks in `references\u002Fhardware-container.md`;\ndefault `nvcr.io\u002Fnvidia\u002Fpytorch:24.09-py3` (driver ≥ 545; SDPA+GQA bug — if\n`num_key_value_heads \u003C num_attention_heads`, set `attn_implementation: \"eager\"`).\nRecord `ngc_image` in `config.yaml`.\n\n**2c. Re-evaluate hardware-dependent compat rules:** re-run the\n`compat-workarounds.md` walk for entries whose `detect` needs `hw`; update\n`applicable_workarounds:` in place.\n\n**2d. Model-fit check:** estimate `param_bytes ≈ 2×param_count` (bf16); if\n> 60% of `vram_gb_per_gpu × 1e9`, recommend LoRA in the user-facing summary.\n\n**Gate:** `config.yaml` has `ngc_image`, `gpu_count`, `gpu_name`, `driver_major`,\n`vram_gb_per_gpu`; hardware-dependent compat fixes recorded.\n\n---\n\n### Step 3 — Research the recipe\n\n**Goal:** fetch the live recipe — training-data knowledge of\n`transformers`\u002F`trl`\u002F`peft` is suspect, so Step 3 is non-negotiable. Walk\n`references\u002Fresearch-priorities.md` in priority order (Priority 1 → 6); stop once\nyou have, for the detected task:\n\n- `AutoModel` \u002F processor class\n- Train + eval transforms\n- Collator\n- `compute_metrics`\n- Hyperparameter hints (LR, batch size, epochs, scheduler)\n\nRecord findings in `meta\u002Frecipe.md`, append source URLs to\n`config.yaml: research_sources:`. A slot with no live finding falls back to the\nmatching scaffold (`cv-scripts.md` \u002F `vlm-scripts.md`), logged as \"fallback to\nscaffold — no live source for \u003Cslot>\" under `notes:`. Conflict-resolution rules\nare in `references\u002Fresearch-priorities.md`.\n\n**Gate:** every required slot filled, with a source URL or scaffold-fallback note.\n\n---\n\n### Step 4 — Generate project & smoke-test\n\n**Goal:** write all scripts, build the image, prepare data, run a 1-step smoke on\nreal data (one `docker build`, two `docker run`s).\n\n**4a. Generate project files** in `output_dir\u002F`: `config.yaml`, `Dockerfile`,\n`requirements.txt`, `prepare_data.py`, `train.py`, `run_eval.py`, `infer.py`,\noptional `merge_lora.py`, optional `tests\u002F`, `.gitignore`. Live Step 3 research is\nauthority; `cv-scripts.md` \u002F `vlm-scripts.md` give scaffold shape only. Apply every\n`applicable_workarounds` entry as a Dockerfile block, requirement pin, config\noverride, or runtime env var. Hard rules: `run_eval.py` keeps that exact filename\n(avoids colliding with the HF `evaluate` package); every generated `.py` starts\nwith the NVIDIA Apache-2.0 copyright header and any emitter fails when it is\nmissing; `emit_unit_tests: true` generates and runs tests per\n`references\u002Ftesting.md`. Script bodies, Dockerfile shape, and the emitter contract\nare in `references\u002Fworkflow-generate-train.md`.\n\n**4b. Build, prepare, smoke** — `docker build -t run-\u003Cshort>:latest .`, then\n`prepare_data` and the `--smoke --max_steps 1` run (`references\u002Fdocker-runs.md`\n§1-3). Smoke pass criteria (in `logs\u002Fsmoke.log`):\n- No exception\n- Loss is finite (not `0.0`, not `NaN`)\n- `grad_norm > 0` at step 1\n\nIf `emit_unit_tests: true`, also run `pytest tests\u002F` in the container. Any failure → STOP.\n\n**4c. Preflight summary** — before full training, print and verify: reference URL,\ndataset columns, Hub target, monitoring target, NGC image, hardware, smoke loss\u002Fgrad norm.\n\n**Gate:** project files written, image built, smoke PASSED, preflight has no\nblank fields.\n\n---\n\n### Step 5 — Train, evaluate, infer\n\n**Goal:** baseline eval, full training, post-train eval, optional LoRA merge, 5\ninference samples (all commands: `references\u002Fdocker-runs.md` §4-8).\n\n| Sub-step | docker-runs.md | Skip if |\n|---|---|---|\n| 5a. Baseline eval (zero-shot) | §4 | `skip_baseline: true` |\n| 5b. Full training (detached) | §5 | — |\n| 5c. LoRA merge | §6 | not VLM+LoRA |\n| 5d. Post-train eval | §7 | — |\n| 5e. Inference (5 samples) | §8 | — |\n\nMulti-GPU: prepend `torchrun --nproc_per_node=$gpu_count` to `python train.py`.\n\nWhile training streams, watch `docker logs -f hft_train`: loss should drop within\n10-20 steps; flat loss (collator\u002Flabel-masking bug), NaN (LR too high), and OOM\nall stop the run — recovery in `references\u002Fcore-rules.md`. If `emit_report: true`,\nrun `report.py` after Step 5e per `references\u002Freporting.md`.\n\n**Gate:** all of:\n- `checkpoints\u002Ffinal\u002F` (or `checkpoints\u002Fmerged\u002F` for LoRA) exists\n- `reports\u002Feval_results.json` has a numeric primary metric\n- `reports\u002Fbaseline_results.json` exists (unless skipped)\n- `reports\u002Finference_samples\u002F` has 5 samples\n- wandb URL shows descending loss\n\n---\n\n### Step 6 — Push & emit rerun skill\n\n**Goal:** publish the run and make it reproducible without re-research.\n\nPush per `references\u002Fhub-push.md` (weights, model card, eval\u002Fbaseline JSONs,\n`config.yaml`, `Dockerfile`, `requirements.txt`, inference samples, reports when\nemitted) unless `push_to_hub: false` is explicit. Emit\n`\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` from\n`references\u002Fpipeline-skill-template.md` — substitute every placeholder, include\nfull YAML metadata + the NVIDIA copyright HTML comment, and make any emitter fail\nif those are missing.\n\n**Gate (Done criteria):** all of:\n- Step 5 gate met\n- HF Hub repo exists at the resolved URL with weights + card + `results\u002F`\n  (unless `push_to_hub: false`)\n- `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` exists, no `\u003Cplaceholder>` left,\n  with metadata + copyright HTML comment per `pipeline-skill-template.md`\n\nFinal message: wandb URL, HF Hub URL, baseline -> fine-tuned primary metric,\n`reports\u002Finference_samples\u002F`, and the rerun skill path.\n\n---\n\n## Error playbook\n\nOn a known runtime error, consult the symptom → minimal-fix table in\n`references\u002Ferror-playbook.md` (NGC entrypoint, PyTorch\u002FTransformers regressions,\nnumpy ABI, Albumentations bbox, PEFT\u002Fcheckpointing, LoRA target breadth, CV\naugmentation gaps, OOM at step 0) before redesigning anything. When a row there\nfires twice across runs, lift it into `compat-workarounds.md` with a `detect` rule\n— auto-applied in Step 1 before the error can fire.\n\n---\n\n## Communication style\n\n- Terse. No filler, no restating the request; one-word answers when appropriate.\n- Always include direct Hub and wandb URLs when referencing artifacts.\n- On error: state what went wrong, why, what you changed — no menus.\n- Never present \"Option A\u002FB\u002FC\" for a request with a clear answer. Act.\n\n## Example pipelines\n\n- [tao-rerun-convnext-cifar10](references\u002Ftao-rerun-convnext-cifar10.md)\n- [tao-rerun-detr-cppe5](references\u002Ftao-rerun-detr-cppe5.md)\n- [tao-rerun-segformer-foodseg103](references\u002Ftao-rerun-segformer-foodseg103.md)\n- [tao-rerun-smolvlm-vqav2](references\u002Ftao-rerun-smolvlm-vqav2.md)\n","\u003C!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0; see http:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0 -->\n\n# tao-finetune-huggingface-model\n\nLocal NVIDIA GPU fine-tuning for HuggingFace models, grounded in live-fetched\ndocumentation with curated references as a fallback safety net. One NGC container,\na few focused scripts, one push to HF Hub. Follow the rules in this file; don't\nimprovise.\n\n**Order of authority (highest first):**\n\n1. **User input** — explicit `model_id`, `dataset_id`, `training_method`, `config.yaml` overrides.\n2. **Live research** — model card, HF repo example, author finetune script, HF task docs, paper; always fetched (Step 3 + `references\u002Fresearch-priorities.md`).\n3. **Curated references** (`references\u002F*.md`) — fallback when live research is silent\u002Fambiguous.\n4. **Your training-data memory** — last resort; suspect, cross-check against (2)\u002F(3).\n\nConflict resolution between (2) and (3) and the source-line discrepancy note are\nin `references\u002Fresearch-priorities.md`.\n\n---\n\n## Inputs\n\n**Required:**\n- `model_id` — HuggingFace model ID, e.g. `google\u002Fvit-base-patch16-224`\n\n**Conditional credentials (read from the session environment, exported before launching when present):**\n- `HF_TOKEN` — only when the model\u002Fdataset is **gated** (read) or `push_to_hub` is on (write); public + public + `push_to_hub: false` needs none. Value never read — presence-only via `[ -n \"$HF_TOKEN\" ]`.\n- `WANDB_API_KEY`, `WANDB_PROJECT` — only when WandB is enabled; `WANDB_MODE=disabled` opts out.\n\n**Dataset — exactly one:**\n- `dataset_id` — HuggingFace dataset ID *(source: `hf`)*\n- `local_dataset_path` — local folder or file *(source: `local`)*; optional\n  `local_dataset_format` ∈ {auto, imagefolder, coco, voc, jsonl, arrow, parquet,\n  csv} (default: auto-detect).\n- *(omit)* — agent recommends popular datasets *(source: `recommend`)*\n\n**Optional (have defaults):**\n- `task_type` — auto-detected from config + model card\n- `n_train=10000`, `n_eval=1000`, `n_epochs=3`, `lora_r=16`\n- `output_dir=.\u002Foutput\u002F\u003Cmodel_short_name>`\n- `hf_model_repo` — push target; if unset and HF_TOKEN has write access,\n  auto-derived as `\u003Cwhoami>\u002F\u003Cmodel_short_name>-finetuned`.\n- `push_to_hub=True` — set to `False` to skip\n- `skip_baseline=False` — skip zero-shot baseline eval\n\n**Optional deliverables (off by default):**\n```yaml\nemit_progress_log: false   # output_dir\u002FPROGRESS.md (per-step journal)\nemit_report:       false   # reports\u002Freport.{pdf,html} with curves & samples\nemit_unit_tests:   false   # tests\u002F with fake-data heterogeneous-batch tests\n```\n\nAll values live in `output_dir\u002Fconfig.yaml`. Never hardcode in Python.\n\n---\n\n## Execution platform\n\nThis skill orchestrates *what* to run; the platform skills own *how* to run it on\na GPU host — read them first.\n\n| Concern | Authoritative skill |\n|---|---|\n| GPU host runtime (driver 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit 1.19.0) | [`tao-skill-bank:tao-setup-nvidia-gpu-host`](..\u002F..\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002FSKILL.md) |\n| `docker run` flags, NGC auth, mounts, env passthrough | [`tao-skill-bank:tao-run-on-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-docker\u002FSKILL.md) |\n| Local Docker job preflight (daemon, GPU smoke) | [`tao-skill-bank:tao-run-on-local-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-local-docker\u002FSKILL.md) |\n\n**Default platform:** `local-docker` — build a one-off image (`run-\u003Cshort>:latest`)\nand run it on the local Docker daemon. Ask only when the user explicitly needs a\ndifferent backend (Brev remote GPU, SLURM\u002FKubernetes); then run that platform's\nPreflight first and route the Steps 4–5 `docker run` commands through it. The\nGPU-runtime and presence-only credential preflights (values never read), the\ncanonical `docker run` flag set, the `list_tao_platforms.py` selection command, and\nthe workflow-specific flags (`--entrypoint \u002Fbin\u002Fbash -lc`, `PYTORCH_CUDA_ALLOC_CONF`,\n`--name hft_train`) are in `references\u002Fworkflow-intake-preflight.md`.\n\n---\n\n## References — fallback safety net\n\nConsulted **only** when live research is silent, ambiguous, or unavailable; live\ndocs always win for the specific model and current API. Each step links the\nreferences it needs; full catalog in `references\u002Fdetailed-workflow.md`.\n\nAlways-on: `core-rules.md`, `error-playbook.md`, `compat-workarounds.md`,\n`model-discovery.md`, `dataset-recommendations.md`, `dataset-sources.md`,\n`dataset-patterns.md`, `hardware-container.md`, `research-priorities.md`,\n`cv-scripts.md`, `vlm-scripts.md`, `docker-runs.md`, `hub-push.md`,\n`pipeline-skill-template.md`, `deliverables.md`. Opt-in (when their flag\u002Fneed\napplies): `progress-tracking.md`, `testing.md`, `reporting.md`,\n`workflow-intake-preflight.md`, `workflow-generate-train.md`, `workflow-push-rerun.md`.\n\n**Rule:** before falling back, log the live source you tried and why it was\ninsufficient (`config.yaml` `notes:`, and PROGRESS.md if enabled). `[FETCH LIVE]`\nmarkers in `cv-scripts.md` \u002F `vlm-scripts.md` are a research checklist, not code to\ninline — refetch the listed URL if a block has no Step 3 finding.\n\n---\n\n## Core rules\n\nNon-negotiable behaviors. **Short version** (full enumeration —\nhallucinated-imports list, never-without-approval list, full error-recovery and\nhardware-sizing tables — in `references\u002Fcore-rules.md`, consult before any\ntraining-time decision):\n\n- **Your HF-library knowledge is outdated.** Fetch live docs (model card, HF\n  repo example, task doc) before writing any ML code — don't generate trainer\n  args \u002F collator \u002F transforms from memory (Step 3).\n- **Smoke-test on real data with `--max_steps 1`** before any full run; no batch\n  launches without a verified smoke.\n- **Never silently substitute** model_id, dataset_id, or training_method — if\n  what the user asked for doesn't load, stop and ask.\n- **Error recovery is minimal-change.** OOM → halve batch, double grad_accum,\n  enable gradient checkpointing (no LoRA switch without approval); NaN → reduce\n  LR 10×; flat loss → inspect collator; same error 3× → stop and ask. Don't loop.\n- **Dataset columns verified BEFORE the collator** — rename in `prepare_data.py`;\n  restructuring needed → stop and ask.\n- **Hardware-sizing thumb (bf16):** ≤3B → 24 GB, 7–13B → 80 GB, 30B+ → multi-GPU\n  or LoRA on 1× 80 GB, 70B+ → 8× 80 GB or LoRA. Full finetune won't fit and no\n  LoRA requested → ask before switching.\n\n---\n\n## Workflow — 6 steps\n\nSingle pass, sequential; each step has a clear gate before the next begins.\n\n### Step 1 — Inspect & qualify\n\n**Goal:** decide whether to proceed. Probe model + dataset, apply accept\u002Freject,\nregister applicable compat fixes, write the initial `config.yaml`.\n\nPrerequisites: `MODEL_ID`, optional `DATASET_ID` \u002F `local_dataset_path`,\noptional `HF_TOKEN`, `OUTPUT_DIR` (default `.\u002Foutput\u002F\u003Cmodel_short_name>`). Probes\nrun in a CPU-only `python:3.12-slim` Docker container (bind-mounted `.probe\u002F`\nscratch) so the host needs no virtualenv — Docker must exist first. Docker-presence\nguard, container env, full probe invocation, and the model\u002Fdataset probe scripts\nare in `references\u002Fworkflow-intake-preflight.md`, `references\u002Fmodel-discovery.md`,\nand `references\u002Fdataset-sources.md`.\n\nProbe requirements:\n\n- Model: load `AutoConfig`, read model-card tags, detect task from\n  `architectures` + tags + card examples (fallback logging in `model-discovery.md`).\n- Dataset: for recommended datasets, first present 3-5 choices from\n  `dataset-recommendations.md`; for local data, bind-mount read-only and use\n  `dataset-sources.md` format detection.\n- Reject early if the model config fails, the task is out of scope, no recipe\n  source exists, or the dataset cannot load \u002F match the task schema.\n- Evaluate `compat-workarounds.md` against the model\u002Ftask; defer hardware-dependent\n  rules to Step 2.\n\nWrite the initial `config.yaml` (`model_id`, `task`, `dataset_id` or\n`local_dataset_path`, `research_sources: []` filled in Step 3,\n`applicable_workarounds:` from Step 1, `notes: []` for reference fallbacks,\n`push_to_hub: true` default — annotated template in\n`references\u002Fworkflow-intake-preflight.md`). Optionally `rm -rf \"$OUTPUT_DIR\u002F.probe\"`\nonce the gate is met.\n\n**Gate:** `config.yaml` exists with model, dataset, task, applicable_workarounds;\ndo not proceed if any field is missing.\n\n---\n\n### Step 2 — Hardware audit & NGC image\n\n**Goal:** verify Docker + GPU + disk, pick the NGC PyTorch image live, finalize\nhardware-dependent compat rules.\n\n**2a. Audit (hard gate)** — three checks (commands in\n`references\u002Fworkflow-intake-preflight.md`):\n1. GPU host runtime — `tao-setup-nvidia-gpu-host`'s\n   `setup-nvidia-gpu-host.sh --backend docker --check-only`; on fail, ask approval\n   then re-run with `--install --yes`.\n2. Free-disk soft-warn — override via `MIN_DISK_GB` (default 100 GB); recommend\n   ≥ 100 GB for NGC base (~20 GB) + HF cache + checkpoints + data.\n3. Conditional credential presence (from the session environment, values never\n   read) — `HF_TOKEN` only when gated or `push_to_hub` is on; `WANDB_*` only when\n   WandB is on.\n\n**Do not proceed to Step 4 on a hard-fail** — Step 4's `docker build` pulls a\n20+ GB NGC base, and a missing `nvidia-container-toolkit` only surfaces later as\n`could not select device driver \"\" with capabilities: [[gpu]]`. Record `gpu_count`,\n`gpu_name`, `driver_major`, `vram_gb_per_gpu` in `config.yaml`.\n\n**2b. Pick NGC image (live):** from the NVIDIA Deep Learning Frameworks support\nmatrix (\u003Chttps:\u002F\u002Fdocs.nvidia.com\u002Fdeeplearning\u002Fframeworks\u002Fsupport-matrix\u002Findex.html>),\nPyTorch NGC container section, pick the highest-versioned image where\n`Min driver ≤ detected driver_major` and container CUDA `≤` host CUDA Toolkit\n(match closely so cuDNN \u002F TensorRT line up). Do **not** reject an image for an\n`aN`\u002F`bN`\u002F`rcN` PyTorch tag — NGC validates the full image; pick the newest\nCUDA-aligned one and let `compat-workarounds.md` handle per-version issues. If the\nmatrix is unreachable, use the fallbacks in `references\u002Fhardware-container.md`;\ndefault `nvcr.io\u002Fnvidia\u002Fpytorch:24.09-py3` (driver ≥ 545; SDPA+GQA bug — if\n`num_key_value_heads \u003C num_attention_heads`, set `attn_implementation: \"eager\"`).\nRecord `ngc_image` in `config.yaml`.\n\n**2c. Re-evaluate hardware-dependent compat rules:** re-run the\n`compat-workarounds.md` walk for entries whose `detect` needs `hw`; update\n`applicable_workarounds:` in place.\n\n**2d. Model-fit check:** estimate `param_bytes ≈ 2×param_count` (bf16); if\n> 60% of `vram_gb_per_gpu × 1e9`, recommend LoRA in the user-facing summary.\n\n**Gate:** `config.yaml` has `ngc_image`, `gpu_count`, `gpu_name`, `driver_major`,\n`vram_gb_per_gpu`; hardware-dependent compat fixes recorded.\n\n---\n\n### Step 3 — Research the recipe\n\n**Goal:** fetch the live recipe — training-data knowledge of\n`transformers`\u002F`trl`\u002F`peft` is suspect, so Step 3 is non-negotiable. Walk\n`references\u002Fresearch-priorities.md` in priority order (Priority 1 → 6); stop once\nyou have, for the detected task:\n\n- `AutoModel` \u002F processor class\n- Train + eval transforms\n- Collator\n- `compute_metrics`\n- Hyperparameter hints (LR, batch size, epochs, scheduler)\n\nRecord findings in `meta\u002Frecipe.md`, append source URLs to\n`config.yaml: research_sources:`. A slot with no live finding falls back to the\nmatching scaffold (`cv-scripts.md` \u002F `vlm-scripts.md`), logged as \"fallback to\nscaffold — no live source for \u003Cslot>\" under `notes:`. Conflict-resolution rules\nare in `references\u002Fresearch-priorities.md`.\n\n**Gate:** every required slot filled, with a source URL or scaffold-fallback note.\n\n---\n\n### Step 4 — Generate project & smoke-test\n\n**Goal:** write all scripts, build the image, prepare data, run a 1-step smoke on\nreal data (one `docker build`, two `docker run`s).\n\n**4a. Generate project files** in `output_dir\u002F`: `config.yaml`, `Dockerfile`,\n`requirements.txt`, `prepare_data.py`, `train.py`, `run_eval.py`, `infer.py`,\noptional `merge_lora.py`, optional `tests\u002F`, `.gitignore`. Live Step 3 research is\nauthority; `cv-scripts.md` \u002F `vlm-scripts.md` give scaffold shape only. Apply every\n`applicable_workarounds` entry as a Dockerfile block, requirement pin, config\noverride, or runtime env var. Hard rules: `run_eval.py` keeps that exact filename\n(avoids colliding with the HF `evaluate` package); every generated `.py` starts\nwith the NVIDIA Apache-2.0 copyright header and any emitter fails when it is\nmissing; `emit_unit_tests: true` generates and runs tests per\n`references\u002Ftesting.md`. Script bodies, Dockerfile shape, and the emitter contract\nare in `references\u002Fworkflow-generate-train.md`.\n\n**4b. Build, prepare, smoke** — `docker build -t run-\u003Cshort>:latest .`, then\n`prepare_data` and the `--smoke --max_steps 1` run (`references\u002Fdocker-runs.md`\n§1-3). Smoke pass criteria (in `logs\u002Fsmoke.log`):\n- No exception\n- Loss is finite (not `0.0`, not `NaN`)\n- `grad_norm > 0` at step 1\n\nIf `emit_unit_tests: true`, also run `pytest tests\u002F` in the container. Any failure → STOP.\n\n**4c. Preflight summary** — before full training, print and verify: reference URL,\ndataset columns, Hub target, monitoring target, NGC image, hardware, smoke loss\u002Fgrad norm.\n\n**Gate:** project files written, image built, smoke PASSED, preflight has no\nblank fields.\n\n---\n\n### Step 5 — Train, evaluate, infer\n\n**Goal:** baseline eval, full training, post-train eval, optional LoRA merge, 5\ninference samples (all commands: `references\u002Fdocker-runs.md` §4-8).\n\n| Sub-step | docker-runs.md | Skip if |\n|---|---|---|\n| 5a. Baseline eval (zero-shot) | §4 | `skip_baseline: true` |\n| 5b. Full training (detached) | §5 | — |\n| 5c. LoRA merge | §6 | not VLM+LoRA |\n| 5d. Post-train eval | §7 | — |\n| 5e. Inference (5 samples) | §8 | — |\n\nMulti-GPU: prepend `torchrun --nproc_per_node=$gpu_count` to `python train.py`.\n\nWhile training streams, watch `docker logs -f hft_train`: loss should drop within\n10-20 steps; flat loss (collator\u002Flabel-masking bug), NaN (LR too high), and OOM\nall stop the run — recovery in `references\u002Fcore-rules.md`. If `emit_report: true`,\nrun `report.py` after Step 5e per `references\u002Freporting.md`.\n\n**Gate:** all of:\n- `checkpoints\u002Ffinal\u002F` (or `checkpoints\u002Fmerged\u002F` for LoRA) exists\n- `reports\u002Feval_results.json` has a numeric primary metric\n- `reports\u002Fbaseline_results.json` exists (unless skipped)\n- `reports\u002Finference_samples\u002F` has 5 samples\n- wandb URL shows descending loss\n\n---\n\n### Step 6 — Push & emit rerun skill\n\n**Goal:** publish the run and make it reproducible without re-research.\n\nPush per `references\u002Fhub-push.md` (weights, model card, eval\u002Fbaseline JSONs,\n`config.yaml`, `Dockerfile`, `requirements.txt`, inference samples, reports when\nemitted) unless `push_to_hub: false` is explicit. Emit\n`\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` from\n`references\u002Fpipeline-skill-template.md` — substitute every placeholder, include\nfull YAML metadata + the NVIDIA copyright HTML comment, and make any emitter fail\nif those are missing.\n\n**Gate (Done criteria):** all of:\n- Step 5 gate met\n- HF Hub repo exists at the resolved URL with weights + card + `results\u002F`\n  (unless `push_to_hub: false`)\n- `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` exists, no `\u003Cplaceholder>` left,\n  with metadata + copyright HTML comment per `pipeline-skill-template.md`\n\nFinal message: wandb URL, HF Hub URL, baseline -> fine-tuned primary metric,\n`reports\u002Finference_samples\u002F`, and the rerun skill path.\n\n---\n\n## Error playbook\n\nOn a known runtime error, consult the symptom → minimal-fix table in\n`references\u002Ferror-playbook.md` (NGC entrypoint, PyTorch\u002FTransformers regressions,\nnumpy ABI, Albumentations bbox, PEFT\u002Fcheckpointing, LoRA target breadth, CV\naugmentation gaps, OOM at step 0) before redesigning anything. When a row there\nfires twice across runs, lift it into `compat-workarounds.md` with a `detect` rule\n— auto-applied in Step 1 before the error can fire.\n\n---\n\n## Communication style\n\n- Terse. No filler, no restating the request; one-word answers when appropriate.\n- Always include direct Hub and wandb URLs when referencing artifacts.\n- On error: state what went wrong, why, what you changed — no menus.\n- Never present \"Option A\u002FB\u002FC\" for a request with a clear answer. Act.\n\n## Example pipelines\n\n- [tao-rerun-convnext-cifar10](references\u002Ftao-rerun-convnext-cifar10.md)\n- [tao-rerun-detr-cppe5](references\u002Ftao-rerun-detr-cppe5.md)\n- [tao-rerun-segformer-foodseg103](references\u002Ftao-rerun-segformer-foodseg103.md)\n- [tao-rerun-smolvlm-vqav2](references\u002Ftao-rerun-smolvlm-vqav2.md)\n",{"name":5,"tags":24,"license":15,"metadata":30,"description":33,"allowed-tools":34,"compatibility":35},[25,26,27,28,29],"finetuning","huggingface","nvidia-tao","computer-vision","training",{"author":31,"version":32},"NVIDIA Corporation","0.1.0","Fine-tune any HuggingFace CV \u002F VLM \u002F LLM model on local NVIDIA GPUs inside an NGC PyTorch container. Use when the user wants to fine-tune a HuggingFace model (full or LoRA), train a vision \u002F VLM \u002F LLM model end-to-end, generate a reproducible HF training pipeline, smoke-test a HuggingFace model locally before scale-up, push a fine-tuned model to the HF Hub with a model card, or emit a self-contained rerun skill for an existing HuggingFace finetune. Supports image classification, object detection, semantic \u002F instance \u002F panoptic segmentation, depth estimation, image-text-to-text VLM (SFT \u002F LoRA), and LLM SFT \u002F DPO \u002F GRPO. Six-step workflow: inspect and qualify, hardware and NGC image, research, generate and smoke, train + eval + infer, push and emit rerun skill.\n","Read Bash Write","Requires docker + nvidia-container-toolkit, NVIDIA GPU (driver ≥ 545, ≥ 24 GB VRAM for ≤3B models), ~40 GB free disk. Optional credentials (read from the session environment, exported before launching) — HF_TOKEN is read only when the model\u002Fdataset is gated or `push_to_hub` is on; WANDB_API_KEY and WANDB_PROJECT only when WandB logging is enabled.","main","8859218955d9504cd18b59e7892100718039d3b4","data\u002Fskills\u002Fnvidia-skills\u002Ftao-finetune-huggingface-model\u002Fskills\u002Ftao-finetune-huggingface-model",1,"https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper","https:\u002F\u002Fgithub.com\u002Fom-scogo","SDE-1 @ Scogo ","2026-07-18 12:22:46",{"source":8,"skillName":5,"commands":45,"prompt":55,"zipAvailable":56},[46,49,52],{"tool":47,"command":48},"npx","npx skills add om-scogo\u002Fskillsh-scraper --skill tao-finetune-huggingface-model",{"tool":50,"command":51},"bunx","bunx skills add om-scogo\u002Fskillsh-scraper --skill tao-finetune-huggingface-model",{"tool":53,"command":54},"pnpm","pnpm dlx skills add om-scogo\u002Fskillsh-scraper --skill tao-finetune-huggingface-model","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fnvidia-skills\u002Ftao-finetune-huggingface-model\u002Fskills\u002Ftao-finetune-huggingface-model\u002FSKILL.md",true,[],[59,67,72,79,84,91],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":63,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7D6MC9FAAEV4CMNYX8W7JTG","arxiv-search","Search arXiv physics, math, and computer science preprints using natural language queries. Powered by Valyu semantic search.","MIT","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fyorkeccak-scientific-skills\u002Farxiv-search\u002Fskills\u002Farxiv-search\u002FSKILL.md",[],[],{"skillNo":68,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":63,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M7D6M85CKKN87TA98VCQV4H7","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fyorkeccak-scientific-skills\u002Farxiv-search\u002FSkill.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":63,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":19},"01M7D4ZTCS4DR3CE4JY06X84T2","working-with-legacy-code","Safely change and test untested codebases using Feathers' \"Working Effectively with Legacy Code\". Use when the user mentions \"legacy code\", \"no tests\", \"untested codebase\", \"how do I test this\", \"seams\", \"characterization tests\", \"golden master\", \"sprout method\", \"afraid to change this code\", \"monster method\", \"dependency breaking\", or \"inherited a messy codebase\". Also trigger when changing code without tests safely, getting a class under test when constructors, statics, or singletons block it, adding features to tangled modules, or planning incremental test coverage for an old codebase. Covers the legacy-code change algorithm, seams, characterization tests, sprout\u002Fwrap, and dependency-breaking techniques. For refactoring code that already has tests, see refactoring-patterns. For day-to-day code quality, see clean-code.","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fworking-with-legacy-code\u002Fworking-with-legacy-code\u002FSKILL.md",[],[],{"skillNo":80,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":63,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M7D4ZHZKJD2RSNGAF63V14P4","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fworking-with-legacy-code\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":63,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":19},"01M7D4ZB4CJBFW5JE67CPH2QRV","web-typography","Select, pair, and implement typefaces for web projects. Use when the user mentions \"font pairing\", \"which typeface\", \"line height\", \"responsive typography\", \"web font loading\", \"type hierarchy\", \"variable fonts\", \"FOUT\u002FFOIT\", \"typographic scale\", or \"the text is hard to read\". Also trigger when choosing between system fonts and web fonts, optimizing font-loading performance, or designing readable long-form content. Covers readability evaluation, CSS implementation, and performance optimization. For overall UI design systems, see refactoring-ui. For dramatic typographic experiences, see top-design.","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fweb-typography\u002Fweb-typography\u002FSKILL.md",[],[],{"skillNo":92,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":63,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M7D4Z4MVYVY3NMJ18BNV6WNA","https:\u002F\u002Fgithub.com\u002Fom-scogo\u002Fskillsh-scraper\u002Fblob\u002Fmain\u002Fdata\u002Fskills\u002Fwondelai-skills\u002Fweb-typography\u002FSkill.md",[],[]]