[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79FPQVRXW592FV20GPZQNWF":3,"skill-same:01M79FPQVRXW592FV20GPZQNWF":349},{"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":14,"repositoryHtmlUrl":39,"repositoryHomepageUrl":6,"ownerHtmlUrl":40,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":41,"install":42,"resources":55},"01M79FPQVRXW592FV20GPZQNWF","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.","Elogic360\u002Fsoftware-factory","Elogic360","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F156535519?v=4","01M74NH3F0RB13PWB7JZDE2V80","elogic360","01M74NH3FPC65082MDZVMFHDM3",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\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 (loaded from `~\u002F.config\u002Ftao\u002F.env` by the SessionStart hook) — 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 WebFetch\n---\n\u003C!--\nCopyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n    http:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License.\n-->\n\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\ncontainer, a small set of focused scripts, one push to HF Hub. Behavior is\ngoverned by the rules in this file — follow them, do not improvise.\n\n**Order of authority (highest first):** (1) user input → (2) live research\n(model card, HF repo example, author script, task docs, paper — always fetched,\nStep 3) → (3) curated `references\u002F*.md` (fallback when live research is silent) →\n(4) training-data memory (last resort, suspect). On conflict, live research wins\nfor the specific model + current API. See `references\u002Fcore-rules.md` for the\nfull order and conflict-resolution rules.\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 (loaded by the SessionStart hook from `~\u002F.config\u002Ftao\u002F.env`):**\n- `HF_TOKEN` — only when the model\u002Fdataset is **gated** (read) or `push_to_hub` is on (write); public + `push_to_hub: false` runs don't need it. The agent never reads the value — only checks presence with `[ -n \"$HF_TOKEN\" ]`.\n- `WANDB_API_KEY`, `WANDB_PROJECT` — only when WandB is enabled; set `WANDB_MODE=disabled` to opt 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 `local_dataset_format` ∈ {auto, imagefolder, coco, voc, jsonl, arrow, parquet, csv} (default auto-detect).\n- *(omit)* — agent recommends popular datasets *(source: `recommend`)*\n\n**Optional (have defaults):** `task_type` (auto-detected); `n_train=10000`,\n`n_eval=1000`, `n_epochs=3`, `lora_r=16`; `output_dir=.\u002Foutput\u002F\u003Cmodel_short_name>`;\n`hf_model_repo` (push target; if unset and HF_TOKEN has write access,\nauto-derived as `\u003Cwhoami>\u002F\u003Cmodel_short_name>-finetuned`); `push_to_hub=True`\n(set `False` to skip); `skip_baseline=False` (skip zero-shot baseline eval).\n\n**Optional deliverables (off by default):** `emit_progress_log` →\n`output_dir\u002FPROGRESS.md` (per-step ✅\u002F⚠️\u002F❌ journal); `emit_report` →\n`reports\u002Freport.{pdf,html}` with curves & samples; `emit_unit_tests` →\n`tests\u002F` with fake-data heterogeneous-batch tests.\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* (read them\nfirst, do not redraft their conventions here):\n[`tao-setup-nvidia-gpu-host`](..\u002F..\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002FSKILL.md)\n(GPU host runtime — driver 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit\n1.19.0), [`tao-run-on-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-docker\u002FSKILL.md)\n(`docker run` flags, NGC auth, `--gpus`, mounts, env passthrough,\n`--ipc=host`\u002F`--shm-size`, error modes), and\n[`tao-run-on-local-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-local-docker\u002FSKILL.md)\n(local Docker job preflight — daemon reachable, GPU smoke).\n\n**Default platform:** `local-docker` — build a one-off image\n(`run-\u003Cshort>:latest`) and run it on the local Docker daemon. Ask only if the\nuser needs a different backend (Brev, Lepton\u002FSLURM\u002FKubernetes). See\n`references\u002Fexecution-platform.md` for that path plus the alternate-backend\nrouting, the GPU-runtime preflight, the credentials policy, and the `docker run`\nconventions.\n\n---\n\n## References — fallback safety net\n\nCurated `references\u002F*.md` are consulted **only** when live research is silent,\nambiguous, or unavailable; live docs always win for the specific model + current\nAPI. The workflow steps below link the file each step needs directly. Before\nfalling back, log the live source you tried and why it was insufficient (in\n`config.yaml` `notes:`, and PROGRESS.md if enabled). `[FETCH LIVE]` markers in\n`cv-scripts.md` \u002F `vlm-scripts.md` are a research checklist, not code to inline —\nif a block has no Step 3 finding, refetch the listed URL.\n\nSee `references\u002Freference-index.md` for the complete index — every always-on\nreference plus the three opt-in ones gated by a flag (`progress-tracking.md` ←\n`emit_progress_log`, `testing.md` ← `emit_unit_tests`, `reporting.md` ←\n`emit_report`), each with its per-step role.\n\n---\n\n## Core rules\n\nThe non-negotiable behaviors. Full text in `references\u002Fcore-rules.md`.\n**Short version:**\n\n- **Your HF-library knowledge is outdated.** Fetch live docs before writing any\n  ML code; never generate trainer args \u002F collator \u002F transforms from memory (Step 3).\n- **Smoke-test on real data with `--max_steps 1`** before any full run.\n- **Never silently substitute** model_id, dataset_id, or training_method — stop and ask.\n- **Error recovery is minimal-change.** OOM → halve batch, double grad_accum,\n  enable gradient checkpointing (don't switch to LoRA without approval); NaN →\n  reduce LR 10×; flat loss → inspect collator; same error 3× → stop and ask.\n- **Dataset columns verified BEFORE the collator.** Rename → `prepare_data.py`;\n  restructuring → stop and ask.\n- **Hardware sizing (bf16):** ≤3B → 24 GB, 7–13B → 80 GB, 30B+ → multi-GPU or\n  LoRA on 1× 80 GB, 70B+ → 8× 80 GB or LoRA. Won't fit + no LoRA request → ask.\n\n`references\u002Fcore-rules.md` has the full enumeration (hallucinated imports,\nnever-without-approval list, full error-recovery + hardware-sizing tables).\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\nDecide whether to proceed at all. **1a. Probe model** and **1b. Probe dataset**\nvia two CPU-only `python:3.12-slim` containerized probes (no host Python\nprereqs): the model probe reports `model_type`, `architectures`, `tags`, head\ncounts; the dataset probe verifies loadability + column schema. Detect `task`\nfrom `architectures` + `tags` + card body (card silent on\n`AutoModelFor...` → `references\u002Fmodel-discovery.md`, log under `notes:`). For\n`source = recommend`, present 3–5 picks from\n`references\u002Fdataset-recommendations.md`; for `source = local`, use\n`references\u002Fdataset-sources.md` loaders. **1c. Accept\u002Freject**, **1d. walk\n`references\u002Fcompat-workarounds.md`** recording matches in `config.yaml`\n`applicable_workarounds:`, then **1e. write the `config.yaml` skeleton**.\n\nSee `references\u002Fstep1-probes.md` for the full probe scripts + `docker run`\ninvocations, the Docker-daemon preflight, prerequisites (`MODEL_ID`, optional\n`DATASET_ID`\u002F`HF_TOKEN`, `OUTPUT_DIR` default `.\u002Foutput\u002F\u003Cmodel_short_name>`\nbind-mounted by Steps 4–5), dataset-column verification + rename rule, the full\nreject criteria, compat-walk detail, the exact skeleton, and `.probe` cleanup.\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\nVerify Docker + GPU + disk, pick the NGC PyTorch image live, finalize\nhardware-dependent compat rules. **2a. Audit (hard gate)** via\n`tao-setup-nvidia-gpu-host --check-only` (driver branch 580, CUDA Toolkit 13.0,\nNVIDIA Container Toolkit 1.19.0); on failure ask to authorize the install, then\nre-run; soft-warn on `\u003C 100 GB` free disk; check only the credentials this run\nneeds; **do not proceed to Step 4 on a hard-fail**; record `gpu_count`,\n`gpu_name`, `driver_major`, `vram_gb_per_gpu`. **2b. Pick NGC image (live)** —\nhighest-versioned PyTorch NGC image with `Min driver ≤ driver_major` and\ncontainer CUDA `≤` host CUDA Toolkit (never reject for an `aN`\u002F`bN`\u002F`rcN`\nsuffix); WebFetch fail → `references\u002Fhardware-container.md` fallback. **2c.\nRe-evaluate** `hw`-dependent compat rules. **2d. Model-fit check** — bf16\n`param_bytes ≈ 2×param_count`; if > 60% of `vram_gb_per_gpu × 1e9`, recommend\nLoRA.\n\nSee `references\u002Fhardware-audit-ngc.md` for the full audit script, the soft-warn\n+ `MIN_DISK_GB` override, live-selection rules, the support-matrix WebFetch URL,\nthe `24.09-py3` \u002F SDPA+GQA `attn_implementation: \"eager\"` fallback, and the\n`could not select device driver` failure note.\n\n**Gate:** `config.yaml` has `ngc_image`, `gpu_count`, `gpu_name`, `driver_major`,\n`vram_gb_per_gpu`. Hardware-dependent compat fixes are recorded.\n\n---\n\n### Step 3 — Research the recipe\n\nFetch the live recipe — the agent's `transformers`\u002F`trl`\u002F`peft` memory is\nsuspect, so Step 3 is non-negotiable. Walk `references\u002Fresearch-priorities.md`\nin priority order (Priority 1 → 6).\nStop once you have, for the detected task: the `AutoModel` \u002F processor class,\ntrain + eval transforms, collator, `compute_metrics`, and hyperparameter hints\n(LR, batch size, epochs, scheduler). Record findings in `meta\u002Frecipe.md` and\nappend source URLs to `config.yaml: research_sources:`. If a slot has no live\nfinding, fall back to the matching scaffold (`cv-scripts.md` \u002F\n`vlm-scripts.md`) and log \"fallback to scaffold — no live source for \u003Cslot>\"\nunder `notes:`. Conflict-resolution rules: `references\u002Fresearch-priorities.md`.\n\n**Gate:** every required slot above is filled, with a source URL or an explicit\nscaffold-fallback note.\n\n---\n\n### Step 4 — Generate project & smoke-test\n\nWrite all scripts, build the image, prepare data, run a 1-step smoke on real\ndata (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` (eval script\n**MUST** be `run_eval.py`, never `evaluate.py` — collides with HF `evaluate`),\n`infer.py`, `merge_lora.py` for VLM-LoRA, `.gitignore`. Authority order: Step 3\nlive research → scaffold reference (`cv-scripts.md` \u002F `vlm-scripts.md`) for\n**structure only**, never their `[FETCH LIVE]` blocks. Apply each\n`applicable_workarounds` entry as a Dockerfile block, requirements pin, config\noverride, or runtime env var. Every generated `.py` begins with the NVIDIA\nApache-2.0 `#`-comment copyright header (emitter must fail otherwise). If\n`emit_unit_tests: true`, also generate `tests\u002F` per `references\u002Ftesting.md`. See\n`references\u002Fproject-scaffold.md` for the full file table, the exact copyright\nheader, and the Dockerfile template (deps → compat → code layer order).\n\n**4b. Build, prepare, smoke** — `docker build -t run-\u003Cshort>:latest .`, then run\n`references\u002Fdocker-runs.md` §1 (build), §2 (prepare_data), §3 (smoke,\n`--smoke --max_steps 1`); §3 lists the smoke pass criteria (no exception, loss\nfinite, `grad_norm > 0` at step 1). If `emit_unit_tests: true`, also run\n`pytest tests\u002F` inside the container. Any failure → STOP.\n\n**4c. Preflight summary** — print the boxed `─ PREFLIGHT ─` summary (reference\nURL, dataset columns, push_to_hub repo, wandb monitoring, ngc_image, hardware,\nsmoke result) and verify every field is filled before launching full training.\nExact format: `references\u002Fproject-scaffold.md`.\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\nRun in order, all commands in `references\u002Fdocker-runs.md`: **5a** baseline eval\n(§4, skip if `skip_baseline: true`), **5b** full training detached (§5), **5c**\nLoRA merge (§6, only VLM-with-LoRA), **5d** post-train eval (§7), **5e**\ninference 5 samples (§8). Multi-GPU: prepend `torchrun --nproc_per_node=$gpu_count`\nto `python train.py`. Watch `docker logs -f hft_train`: loss should drop within\n10-20 steps (flat → stop; NaN → reduce LR; OOM → halve batch; full recovery in\n`references\u002Fcore-rules.md` + `references\u002Ferror-playbook.md`). If\n`emit_report: true`, run `report.py` after Step 5e per `references\u002Freporting.md`.\n\n**Gate:** all of — `checkpoints\u002Ffinal\u002F` (or `checkpoints\u002Fmerged\u002F` for LoRA)\nexists; `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; wandb URL shows descending loss.\n\n---\n\n### Step 6 — Push & emit rerun skill\n\nPublish the run and make it reproducible without re-research.\n\n**6a. Push to HF Hub** — use `references\u002Fhub-push.md` (pushes weights merged or\nfinal, a generated model card `README.md`, `results\u002F{eval,baseline}_results.json`,\n`config.yaml`, `Dockerfile`, `requirements.txt`, `inference_samples\u002F*.jpg`, and\n`report.{pdf,html}` if `emit_report: true`). Skip iff `push_to_hub: false` is\nexplicit in `config.yaml`.\n\n**6b. Emit rerun skill** at `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` per\n`references\u002Fpipeline-skill-template.md`. Every `\u003Cplaceholder>` must be a real\nvalue (literal placeholders are a bug); include the full YAML (`license`,\n`compatibility`, `metadata`, `allowed-tools`) and the NVIDIA copyright notice in\nan HTML comment immediately after the closing `---`, as in that template; an\nemitter must fail unless the emitted `SKILL.md` contains those fields and the\ncopyright comment.\n\n**Gate (Done criteria):** all of — Step 5 gate met; HF Hub repo exists at the\nresolved URL with weights + card + `results\u002F` (unless `push_to_hub: false`);\n`\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` exists with no `\u003Cplaceholder>` left,\nwith metadata + copyright HTML comment per `pipeline-skill-template.md`.\n\n**Final message to user** — terse, with direct URLs: wandb URL; HF Hub URL;\nprimary metric baseline → fine-tuned (Δ); path to `reports\u002Finference_samples\u002F`;\npath to `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md`.\n\n---\n\n## Error playbook\n\nOn a known runtime error, consult `references\u002Ferror-playbook.md` before\nredesigning anything — its symptom → minimal-fix table covers NGC ENTRYPOINT,\nSDPA+GQA, `transformers>=4.51` regression, numpy 2.x ABI, Albumentations bbox,\nPEFT + gradient_checkpointing, SmolVLM SDPA, LoRA target-regex, missing CV\naugmentation, OOM at step 0, and more. When a row fires twice across runs, lift\nit into `references\u002Fcompat-workarounds.md` with a `detect` rule, auto-applied in\nStep 1d before the error can fire.\n\n---\n\n## Communication style\n\nTerse: no filler, no restating the request; always include direct Hub + wandb\nURLs; on error state what went wrong, why, what you changed (no menus, no\n\"Option A\u002FB\u002FC\" when the answer is clear — act). Full text:\n`references\u002Fcore-rules.md`.\n\n## Example pipelines\n\n- [tao-rerun-convnext-cifar10](references\u002Ftao-rerun-convnext-cifar10.md) — facebook\u002Fconvnext-tiny-224 on cifar10 (image-classification, 10 classes, subset 5000\u002F1000).\n- [tao-rerun-detr-cppe5](references\u002Ftao-rerun-detr-cppe5.md) — facebook\u002Fdetr-resnet-50 on cppe-5 (object-detection, 5 classes, subset 800\u002F200).\n- [tao-rerun-segformer-foodseg103](references\u002Ftao-rerun-segformer-foodseg103.md) — nvidia\u002Fmit-b0 on EduardoPacheco\u002FFoodSeg103 (semantic segmentation, 103 classes + background, subset 1000\u002F200).\n- [tao-rerun-smolvlm-vqav2](references\u002Ftao-rerun-smolvlm-vqav2.md) — HuggingFaceTB\u002FSmolVLM-256M-Instruct on merve\u002Fvqav2-small (image-text-to-text VLM LoRA, subset 500\u002F100, 5 epochs).\n","\u003C!--\nCopyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.\n\nLicensed under the Apache License, Version 2.0 (the \"License\");\nyou may not use this file except in compliance with the License.\nYou may obtain a copy of the License at\n\n    http:\u002F\u002Fwww.apache.org\u002Flicenses\u002FLICENSE-2.0\n\nUnless required by applicable law or agreed to in writing, software\ndistributed under the License is distributed on an \"AS IS\" BASIS,\nWITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\nSee the License for the specific language governing permissions and\nlimitations under the License.\n-->\n\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\ncontainer, a small set of focused scripts, one push to HF Hub. Behavior is\ngoverned by the rules in this file — follow them, do not improvise.\n\n**Order of authority (highest first):** (1) user input → (2) live research\n(model card, HF repo example, author script, task docs, paper — always fetched,\nStep 3) → (3) curated `references\u002F*.md` (fallback when live research is silent) →\n(4) training-data memory (last resort, suspect). On conflict, live research wins\nfor the specific model + current API. See `references\u002Fcore-rules.md` for the\nfull order and conflict-resolution rules.\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 (loaded by the SessionStart hook from `~\u002F.config\u002Ftao\u002F.env`):**\n- `HF_TOKEN` — only when the model\u002Fdataset is **gated** (read) or `push_to_hub` is on (write); public + `push_to_hub: false` runs don't need it. The agent never reads the value — only checks presence with `[ -n \"$HF_TOKEN\" ]`.\n- `WANDB_API_KEY`, `WANDB_PROJECT` — only when WandB is enabled; set `WANDB_MODE=disabled` to opt 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 `local_dataset_format` ∈ {auto, imagefolder, coco, voc, jsonl, arrow, parquet, csv} (default auto-detect).\n- *(omit)* — agent recommends popular datasets *(source: `recommend`)*\n\n**Optional (have defaults):** `task_type` (auto-detected); `n_train=10000`,\n`n_eval=1000`, `n_epochs=3`, `lora_r=16`; `output_dir=.\u002Foutput\u002F\u003Cmodel_short_name>`;\n`hf_model_repo` (push target; if unset and HF_TOKEN has write access,\nauto-derived as `\u003Cwhoami>\u002F\u003Cmodel_short_name>-finetuned`); `push_to_hub=True`\n(set `False` to skip); `skip_baseline=False` (skip zero-shot baseline eval).\n\n**Optional deliverables (off by default):** `emit_progress_log` →\n`output_dir\u002FPROGRESS.md` (per-step ✅\u002F⚠️\u002F❌ journal); `emit_report` →\n`reports\u002Freport.{pdf,html}` with curves & samples; `emit_unit_tests` →\n`tests\u002F` with fake-data heterogeneous-batch tests.\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* (read them\nfirst, do not redraft their conventions here):\n[`tao-setup-nvidia-gpu-host`](..\u002F..\u002Fplatform\u002Ftao-setup-nvidia-gpu-host\u002FSKILL.md)\n(GPU host runtime — driver 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit\n1.19.0), [`tao-run-on-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-docker\u002FSKILL.md)\n(`docker run` flags, NGC auth, `--gpus`, mounts, env passthrough,\n`--ipc=host`\u002F`--shm-size`, error modes), and\n[`tao-run-on-local-docker`](..\u002F..\u002Fplatform\u002Ftao-run-on-local-docker\u002FSKILL.md)\n(local Docker job preflight — daemon reachable, GPU smoke).\n\n**Default platform:** `local-docker` — build a one-off image\n(`run-\u003Cshort>:latest`) and run it on the local Docker daemon. Ask only if the\nuser needs a different backend (Brev, Lepton\u002FSLURM\u002FKubernetes). See\n`references\u002Fexecution-platform.md` for that path plus the alternate-backend\nrouting, the GPU-runtime preflight, the credentials policy, and the `docker run`\nconventions.\n\n---\n\n## References — fallback safety net\n\nCurated `references\u002F*.md` are consulted **only** when live research is silent,\nambiguous, or unavailable; live docs always win for the specific model + current\nAPI. The workflow steps below link the file each step needs directly. Before\nfalling back, log the live source you tried and why it was insufficient (in\n`config.yaml` `notes:`, and PROGRESS.md if enabled). `[FETCH LIVE]` markers in\n`cv-scripts.md` \u002F `vlm-scripts.md` are a research checklist, not code to inline —\nif a block has no Step 3 finding, refetch the listed URL.\n\nSee `references\u002Freference-index.md` for the complete index — every always-on\nreference plus the three opt-in ones gated by a flag (`progress-tracking.md` ←\n`emit_progress_log`, `testing.md` ← `emit_unit_tests`, `reporting.md` ←\n`emit_report`), each with its per-step role.\n\n---\n\n## Core rules\n\nThe non-negotiable behaviors. Full text in `references\u002Fcore-rules.md`.\n**Short version:**\n\n- **Your HF-library knowledge is outdated.** Fetch live docs before writing any\n  ML code; never generate trainer args \u002F collator \u002F transforms from memory (Step 3).\n- **Smoke-test on real data with `--max_steps 1`** before any full run.\n- **Never silently substitute** model_id, dataset_id, or training_method — stop and ask.\n- **Error recovery is minimal-change.** OOM → halve batch, double grad_accum,\n  enable gradient checkpointing (don't switch to LoRA without approval); NaN →\n  reduce LR 10×; flat loss → inspect collator; same error 3× → stop and ask.\n- **Dataset columns verified BEFORE the collator.** Rename → `prepare_data.py`;\n  restructuring → stop and ask.\n- **Hardware sizing (bf16):** ≤3B → 24 GB, 7–13B → 80 GB, 30B+ → multi-GPU or\n  LoRA on 1× 80 GB, 70B+ → 8× 80 GB or LoRA. Won't fit + no LoRA request → ask.\n\n`references\u002Fcore-rules.md` has the full enumeration (hallucinated imports,\nnever-without-approval list, full error-recovery + hardware-sizing tables).\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\nDecide whether to proceed at all. **1a. Probe model** and **1b. Probe dataset**\nvia two CPU-only `python:3.12-slim` containerized probes (no host Python\nprereqs): the model probe reports `model_type`, `architectures`, `tags`, head\ncounts; the dataset probe verifies loadability + column schema. Detect `task`\nfrom `architectures` + `tags` + card body (card silent on\n`AutoModelFor...` → `references\u002Fmodel-discovery.md`, log under `notes:`). For\n`source = recommend`, present 3–5 picks from\n`references\u002Fdataset-recommendations.md`; for `source = local`, use\n`references\u002Fdataset-sources.md` loaders. **1c. Accept\u002Freject**, **1d. walk\n`references\u002Fcompat-workarounds.md`** recording matches in `config.yaml`\n`applicable_workarounds:`, then **1e. write the `config.yaml` skeleton**.\n\nSee `references\u002Fstep1-probes.md` for the full probe scripts + `docker run`\ninvocations, the Docker-daemon preflight, prerequisites (`MODEL_ID`, optional\n`DATASET_ID`\u002F`HF_TOKEN`, `OUTPUT_DIR` default `.\u002Foutput\u002F\u003Cmodel_short_name>`\nbind-mounted by Steps 4–5), dataset-column verification + rename rule, the full\nreject criteria, compat-walk detail, the exact skeleton, and `.probe` cleanup.\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\nVerify Docker + GPU + disk, pick the NGC PyTorch image live, finalize\nhardware-dependent compat rules. **2a. Audit (hard gate)** via\n`tao-setup-nvidia-gpu-host --check-only` (driver branch 580, CUDA Toolkit 13.0,\nNVIDIA Container Toolkit 1.19.0); on failure ask to authorize the install, then\nre-run; soft-warn on `\u003C 100 GB` free disk; check only the credentials this run\nneeds; **do not proceed to Step 4 on a hard-fail**; record `gpu_count`,\n`gpu_name`, `driver_major`, `vram_gb_per_gpu`. **2b. Pick NGC image (live)** —\nhighest-versioned PyTorch NGC image with `Min driver ≤ driver_major` and\ncontainer CUDA `≤` host CUDA Toolkit (never reject for an `aN`\u002F`bN`\u002F`rcN`\nsuffix); WebFetch fail → `references\u002Fhardware-container.md` fallback. **2c.\nRe-evaluate** `hw`-dependent compat rules. **2d. Model-fit check** — bf16\n`param_bytes ≈ 2×param_count`; if > 60% of `vram_gb_per_gpu × 1e9`, recommend\nLoRA.\n\nSee `references\u002Fhardware-audit-ngc.md` for the full audit script, the soft-warn\n+ `MIN_DISK_GB` override, live-selection rules, the support-matrix WebFetch URL,\nthe `24.09-py3` \u002F SDPA+GQA `attn_implementation: \"eager\"` fallback, and the\n`could not select device driver` failure note.\n\n**Gate:** `config.yaml` has `ngc_image`, `gpu_count`, `gpu_name`, `driver_major`,\n`vram_gb_per_gpu`. Hardware-dependent compat fixes are recorded.\n\n---\n\n### Step 3 — Research the recipe\n\nFetch the live recipe — the agent's `transformers`\u002F`trl`\u002F`peft` memory is\nsuspect, so Step 3 is non-negotiable. Walk `references\u002Fresearch-priorities.md`\nin priority order (Priority 1 → 6).\nStop once you have, for the detected task: the `AutoModel` \u002F processor class,\ntrain + eval transforms, collator, `compute_metrics`, and hyperparameter hints\n(LR, batch size, epochs, scheduler). Record findings in `meta\u002Frecipe.md` and\nappend source URLs to `config.yaml: research_sources:`. If a slot has no live\nfinding, fall back to the matching scaffold (`cv-scripts.md` \u002F\n`vlm-scripts.md`) and log \"fallback to scaffold — no live source for \u003Cslot>\"\nunder `notes:`. Conflict-resolution rules: `references\u002Fresearch-priorities.md`.\n\n**Gate:** every required slot above is filled, with a source URL or an explicit\nscaffold-fallback note.\n\n---\n\n### Step 4 — Generate project & smoke-test\n\nWrite all scripts, build the image, prepare data, run a 1-step smoke on real\ndata (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` (eval script\n**MUST** be `run_eval.py`, never `evaluate.py` — collides with HF `evaluate`),\n`infer.py`, `merge_lora.py` for VLM-LoRA, `.gitignore`. Authority order: Step 3\nlive research → scaffold reference (`cv-scripts.md` \u002F `vlm-scripts.md`) for\n**structure only**, never their `[FETCH LIVE]` blocks. Apply each\n`applicable_workarounds` entry as a Dockerfile block, requirements pin, config\noverride, or runtime env var. Every generated `.py` begins with the NVIDIA\nApache-2.0 `#`-comment copyright header (emitter must fail otherwise). If\n`emit_unit_tests: true`, also generate `tests\u002F` per `references\u002Ftesting.md`. See\n`references\u002Fproject-scaffold.md` for the full file table, the exact copyright\nheader, and the Dockerfile template (deps → compat → code layer order).\n\n**4b. Build, prepare, smoke** — `docker build -t run-\u003Cshort>:latest .`, then run\n`references\u002Fdocker-runs.md` §1 (build), §2 (prepare_data), §3 (smoke,\n`--smoke --max_steps 1`); §3 lists the smoke pass criteria (no exception, loss\nfinite, `grad_norm > 0` at step 1). If `emit_unit_tests: true`, also run\n`pytest tests\u002F` inside the container. Any failure → STOP.\n\n**4c. Preflight summary** — print the boxed `─ PREFLIGHT ─` summary (reference\nURL, dataset columns, push_to_hub repo, wandb monitoring, ngc_image, hardware,\nsmoke result) and verify every field is filled before launching full training.\nExact format: `references\u002Fproject-scaffold.md`.\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\nRun in order, all commands in `references\u002Fdocker-runs.md`: **5a** baseline eval\n(§4, skip if `skip_baseline: true`), **5b** full training detached (§5), **5c**\nLoRA merge (§6, only VLM-with-LoRA), **5d** post-train eval (§7), **5e**\ninference 5 samples (§8). Multi-GPU: prepend `torchrun --nproc_per_node=$gpu_count`\nto `python train.py`. Watch `docker logs -f hft_train`: loss should drop within\n10-20 steps (flat → stop; NaN → reduce LR; OOM → halve batch; full recovery in\n`references\u002Fcore-rules.md` + `references\u002Ferror-playbook.md`). If\n`emit_report: true`, run `report.py` after Step 5e per `references\u002Freporting.md`.\n\n**Gate:** all of — `checkpoints\u002Ffinal\u002F` (or `checkpoints\u002Fmerged\u002F` for LoRA)\nexists; `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; wandb URL shows descending loss.\n\n---\n\n### Step 6 — Push & emit rerun skill\n\nPublish the run and make it reproducible without re-research.\n\n**6a. Push to HF Hub** — use `references\u002Fhub-push.md` (pushes weights merged or\nfinal, a generated model card `README.md`, `results\u002F{eval,baseline}_results.json`,\n`config.yaml`, `Dockerfile`, `requirements.txt`, `inference_samples\u002F*.jpg`, and\n`report.{pdf,html}` if `emit_report: true`). Skip iff `push_to_hub: false` is\nexplicit in `config.yaml`.\n\n**6b. Emit rerun skill** at `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` per\n`references\u002Fpipeline-skill-template.md`. Every `\u003Cplaceholder>` must be a real\nvalue (literal placeholders are a bug); include the full YAML (`license`,\n`compatibility`, `metadata`, `allowed-tools`) and the NVIDIA copyright notice in\nan HTML comment immediately after the closing `---`, as in that template; an\nemitter must fail unless the emitted `SKILL.md` contains those fields and the\ncopyright comment.\n\n**Gate (Done criteria):** all of — Step 5 gate met; HF Hub repo exists at the\nresolved URL with weights + card + `results\u002F` (unless `push_to_hub: false`);\n`\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md` exists with no `\u003Cplaceholder>` left,\nwith metadata + copyright HTML comment per `pipeline-skill-template.md`.\n\n**Final message to user** — terse, with direct URLs: wandb URL; HF Hub URL;\nprimary metric baseline → fine-tuned (Δ); path to `reports\u002Finference_samples\u002F`;\npath to `\u003Coutput_dir>\u002Fskills\u002Frun-\u003Cshort>\u002FSKILL.md`.\n\n---\n\n## Error playbook\n\nOn a known runtime error, consult `references\u002Ferror-playbook.md` before\nredesigning anything — its symptom → minimal-fix table covers NGC ENTRYPOINT,\nSDPA+GQA, `transformers>=4.51` regression, numpy 2.x ABI, Albumentations bbox,\nPEFT + gradient_checkpointing, SmolVLM SDPA, LoRA target-regex, missing CV\naugmentation, OOM at step 0, and more. When a row fires twice across runs, lift\nit into `references\u002Fcompat-workarounds.md` with a `detect` rule, auto-applied in\nStep 1d before the error can fire.\n\n---\n\n## Communication style\n\nTerse: no filler, no restating the request; always include direct Hub + wandb\nURLs; on error state what went wrong, why, what you changed (no menus, no\n\"Option A\u002FB\u002FC\" when the answer is clear — act). Full text:\n`references\u002Fcore-rules.md`.\n\n## Example pipelines\n\n- [tao-rerun-convnext-cifar10](references\u002Ftao-rerun-convnext-cifar10.md) — facebook\u002Fconvnext-tiny-224 on cifar10 (image-classification, 10 classes, subset 5000\u002F1000).\n- [tao-rerun-detr-cppe5](references\u002Ftao-rerun-detr-cppe5.md) — facebook\u002Fdetr-resnet-50 on cppe-5 (object-detection, 5 classes, subset 800\u002F200).\n- [tao-rerun-segformer-foodseg103](references\u002Ftao-rerun-segformer-foodseg103.md) — nvidia\u002Fmit-b0 on EduardoPacheco\u002FFoodSeg103 (semantic segmentation, 103 classes + background, subset 1000\u002F200).\n- [tao-rerun-smolvlm-vqav2](references\u002Ftao-rerun-smolvlm-vqav2.md) — HuggingFaceTB\u002FSmolVLM-256M-Instruct on merve\u002Fvqav2-small (image-text-to-text VLM LoRA, subset 500\u002F100, 5 epochs).\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 WebFetch","Requires docker + nvidia-container-toolkit, NVIDIA GPU (driver ≥ 545, ≥ 24 GB VRAM for ≤3B models), ~40 GB free disk. Optional credentials (loaded from `~\u002F.config\u002Ftao\u002F.env` by the SessionStart hook) — 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","c475fc64e94e803b50d14a791dfee8853fbb2c5c","skills\u002Fnvidia\u002Ftao-finetune-huggingface-model","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory","https:\u002F\u002Fgithub.com\u002FElogic360","2026-11-08 22:21:22",{"source":8,"skillName":5,"commands":43,"prompt":53,"zipAvailable":54},[44,47,50],{"tool":45,"command":46},"npx","npx skills add Elogic360\u002Fsoftware-factory --skill tao-finetune-huggingface-model",{"tool":48,"command":49},"bunx","bunx skills add Elogic360\u002Fsoftware-factory --skill tao-finetune-huggingface-model",{"tool":51,"command":52},"pnpm","pnpm dlx skills add Elogic360\u002Fsoftware-factory --skill tao-finetune-huggingface-model","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002FSKILL.md",true,[56,62,66,72,76,80,86,90,94,98,102,106,110,113,117,120,124,128,132,136,140,144,148,152,155,159,163,167,171,175,179,183,187,191,194,198,201,205,209,213,217,221,225,229,233,237,241,245,249,253,257,261,265,269,273,277,281,285,289,293,297,301,305,309,313,317,321,325,329,333,337,341,345],{"path":57,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":60,"sourceHtmlUrl":61},"BENCHMARK.md","REFERENCE","参考文档",9332,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002FBENCHMARK.md",{"path":63,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":64,"sourceHtmlUrl":65},"evals\u002Fevals.json",844,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fevals\u002Fevals.json",{"path":67,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":70,"sourceHtmlUrl":71},"examples\u002Fconvnext-tiny-cifar10\u002F.gitignore","OTHER","其他",940,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002F.gitignore",{"path":73,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":74,"sourceHtmlUrl":75},"examples\u002Fconvnext-tiny-cifar10\u002Fconfig.yaml",2110,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Fconfig.yaml",{"path":77,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":78,"sourceHtmlUrl":79},"examples\u002Fconvnext-tiny-cifar10\u002FDockerfile",837,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002FDockerfile",{"path":81,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":84,"sourceHtmlUrl":85},"examples\u002Fconvnext-tiny-cifar10\u002Finfer.py","SCRIPT","脚本",2521,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Finfer.py",{"path":87,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":88,"sourceHtmlUrl":89},"examples\u002Fconvnext-tiny-cifar10\u002Fprepare_data.py",1441,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Fprepare_data.py",{"path":91,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":92,"sourceHtmlUrl":93},"examples\u002Fconvnext-tiny-cifar10\u002Freports\u002Fbaseline_results.json",442,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Freports\u002Fbaseline_results.json",{"path":95,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":96,"sourceHtmlUrl":97},"examples\u002Fconvnext-tiny-cifar10\u002Freports\u002Feval_results.json",422,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Freports\u002Feval_results.json",{"path":99,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":100,"sourceHtmlUrl":101},"examples\u002Fconvnext-tiny-cifar10\u002Frequirements.txt",939,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Frequirements.txt",{"path":103,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":104,"sourceHtmlUrl":105},"examples\u002Fconvnext-tiny-cifar10\u002Frun_eval.py",3087,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Frun_eval.py",{"path":107,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":108,"sourceHtmlUrl":109},"examples\u002Fconvnext-tiny-cifar10\u002Ftrain.py",5611,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fconvnext-tiny-cifar10\u002Ftrain.py",{"path":111,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":70,"sourceHtmlUrl":112},"examples\u002Fdetr-resnet50-cppe5\u002F.gitignore","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002F.gitignore",{"path":114,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":115,"sourceHtmlUrl":116},"examples\u002Fdetr-resnet50-cppe5\u002Fconfig.yaml",2098,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Fconfig.yaml",{"path":118,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":78,"sourceHtmlUrl":119},"examples\u002Fdetr-resnet50-cppe5\u002FDockerfile","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002FDockerfile",{"path":121,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":122,"sourceHtmlUrl":123},"examples\u002Fdetr-resnet50-cppe5\u002Finfer.py",3450,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Finfer.py",{"path":125,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":126,"sourceHtmlUrl":127},"examples\u002Fdetr-resnet50-cppe5\u002Fprepare_data.py",1302,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Fprepare_data.py",{"path":129,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":130,"sourceHtmlUrl":131},"examples\u002Fdetr-resnet50-cppe5\u002Freports\u002Fbaseline_results.json",513,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Freports\u002Fbaseline_results.json",{"path":133,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":134,"sourceHtmlUrl":135},"examples\u002Fdetr-resnet50-cppe5\u002Freports\u002Feval_results.json",488,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Freports\u002Feval_results.json",{"path":137,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":138,"sourceHtmlUrl":139},"examples\u002Fdetr-resnet50-cppe5\u002Frequirements.txt",963,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Frequirements.txt",{"path":141,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":142,"sourceHtmlUrl":143},"examples\u002Fdetr-resnet50-cppe5\u002Frun_eval.py",3703,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Frun_eval.py",{"path":145,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":146,"sourceHtmlUrl":147},"examples\u002Fdetr-resnet50-cppe5\u002Ftrain.py",9243,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fdetr-resnet50-cppe5\u002Ftrain.py",{"path":149,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":150,"sourceHtmlUrl":151},"examples\u002FREADME.md",5939,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002FREADME.md",{"path":153,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":70,"sourceHtmlUrl":154},"examples\u002Fsegformer-b0-foodseg103\u002F.gitignore","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002F.gitignore",{"path":156,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":157,"sourceHtmlUrl":158},"examples\u002Fsegformer-b0-foodseg103\u002Fconfig.yaml",2192,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Fconfig.yaml",{"path":160,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":161,"sourceHtmlUrl":162},"examples\u002Fsegformer-b0-foodseg103\u002FDockerfile",834,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002FDockerfile",{"path":164,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":165,"sourceHtmlUrl":166},"examples\u002Fsegformer-b0-foodseg103\u002Finfer.py",3331,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Finfer.py",{"path":168,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":169,"sourceHtmlUrl":170},"examples\u002Fsegformer-b0-foodseg103\u002Fprepare_data.py",1123,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Fprepare_data.py",{"path":172,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":173,"sourceHtmlUrl":174},"examples\u002Fsegformer-b0-foodseg103\u002Freports\u002Fbaseline_results.json",166,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Freports\u002Fbaseline_results.json",{"path":176,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":177,"sourceHtmlUrl":178},"examples\u002Fsegformer-b0-foodseg103\u002Freports\u002Feval_results.json",168,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Freports\u002Feval_results.json",{"path":180,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":181,"sourceHtmlUrl":182},"examples\u002Fsegformer-b0-foodseg103\u002Frequirements.txt",724,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Frequirements.txt",{"path":184,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":185,"sourceHtmlUrl":186},"examples\u002Fsegformer-b0-foodseg103\u002Frun_eval.py",3129,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Frun_eval.py",{"path":188,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":189,"sourceHtmlUrl":190},"examples\u002Fsegformer-b0-foodseg103\u002Ftrain.py",6742,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsegformer-b0-foodseg103\u002Ftrain.py",{"path":192,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":70,"sourceHtmlUrl":193},"examples\u002Fsmolvlm-256m-vqav2\u002F.gitignore","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002F.gitignore",{"path":195,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":196,"sourceHtmlUrl":197},"examples\u002Fsmolvlm-256m-vqav2\u002Fconfig.yaml",2218,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Fconfig.yaml",{"path":199,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":161,"sourceHtmlUrl":200},"examples\u002Fsmolvlm-256m-vqav2\u002FDockerfile","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002FDockerfile",{"path":202,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":203,"sourceHtmlUrl":204},"examples\u002Fsmolvlm-256m-vqav2\u002Finfer.py",2876,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Finfer.py",{"path":206,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":207,"sourceHtmlUrl":208},"examples\u002Fsmolvlm-256m-vqav2\u002Fmerge_lora.py",1157,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Fmerge_lora.py",{"path":210,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":211,"sourceHtmlUrl":212},"examples\u002Fsmolvlm-256m-vqav2\u002Fprepare_data.py",1156,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Fprepare_data.py",{"path":214,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":215,"sourceHtmlUrl":216},"examples\u002Fsmolvlm-256m-vqav2\u002Freports\u002Fbaseline_results.json",1628,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Freports\u002Fbaseline_results.json",{"path":218,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":219,"sourceHtmlUrl":220},"examples\u002Fsmolvlm-256m-vqav2\u002Freports\u002Feval_results.json",1255,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Freports\u002Feval_results.json",{"path":222,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":223,"sourceHtmlUrl":224},"examples\u002Fsmolvlm-256m-vqav2\u002Frequirements.txt",853,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Frequirements.txt",{"path":226,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":227,"sourceHtmlUrl":228},"examples\u002Fsmolvlm-256m-vqav2\u002Frun_eval.py",2997,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Frun_eval.py",{"path":230,"resourceType":82,"resourceTypeDesc":83,"sizeBytes":231,"sourceHtmlUrl":232},"examples\u002Fsmolvlm-256m-vqav2\u002Ftrain.py",4997,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fexamples\u002Fsmolvlm-256m-vqav2\u002Ftrain.py",{"path":234,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":235,"sourceHtmlUrl":236},"references\u002Fcompat-workarounds.md",11427,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fcompat-workarounds.md",{"path":238,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":239,"sourceHtmlUrl":240},"references\u002Fcore-rules.md",5443,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fcore-rules.md",{"path":242,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":243,"sourceHtmlUrl":244},"references\u002Fcv-scripts.md",33537,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fcv-scripts.md",{"path":246,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":247,"sourceHtmlUrl":248},"references\u002Fdataset-patterns.md",10264,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fdataset-patterns.md",{"path":250,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":251,"sourceHtmlUrl":252},"references\u002Fdataset-recommendations.md",10487,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fdataset-recommendations.md",{"path":254,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":255,"sourceHtmlUrl":256},"references\u002Fdataset-sources.md",16937,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fdataset-sources.md",{"path":258,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":259,"sourceHtmlUrl":260},"references\u002Fdeliverables.md",20001,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fdeliverables.md",{"path":262,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":263,"sourceHtmlUrl":264},"references\u002Fdocker-runs.md",8426,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fdocker-runs.md",{"path":266,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":267,"sourceHtmlUrl":268},"references\u002Ferror-playbook.md",3844,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ferror-playbook.md",{"path":270,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":271,"sourceHtmlUrl":272},"references\u002Fexecution-platform.md",3031,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fexecution-platform.md",{"path":274,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":275,"sourceHtmlUrl":276},"references\u002Fhardware-audit-ngc.md",4399,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fhardware-audit-ngc.md",{"path":278,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":279,"sourceHtmlUrl":280},"references\u002Fhardware-container.md",7954,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fhardware-container.md",{"path":282,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":283,"sourceHtmlUrl":284},"references\u002Fhub-push.md",5222,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fhub-push.md",{"path":286,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":287,"sourceHtmlUrl":288},"references\u002Fmodel-discovery.md",10386,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fmodel-discovery.md",{"path":290,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":291,"sourceHtmlUrl":292},"references\u002Fpipeline-skill-template.md",8743,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fpipeline-skill-template.md",{"path":294,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":295,"sourceHtmlUrl":296},"references\u002Fprogress-tracking.md",9478,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fprogress-tracking.md",{"path":298,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":299,"sourceHtmlUrl":300},"references\u002Fproject-scaffold.md",4309,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fproject-scaffold.md",{"path":302,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":303,"sourceHtmlUrl":304},"references\u002Freference-index.md",3473,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Freference-index.md",{"path":306,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":307,"sourceHtmlUrl":308},"references\u002Freporting.md",18473,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Freporting.md",{"path":310,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":311,"sourceHtmlUrl":312},"references\u002Fresearch-priorities.md",5815,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fresearch-priorities.md",{"path":314,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":315,"sourceHtmlUrl":316},"references\u002Fstep1-probes.md",9457,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fstep1-probes.md",{"path":318,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":319,"sourceHtmlUrl":320},"references\u002Ftao-rerun-convnext-cifar10.md",4571,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ftao-rerun-convnext-cifar10.md",{"path":322,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":323,"sourceHtmlUrl":324},"references\u002Ftao-rerun-detr-cppe5.md",3937,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ftao-rerun-detr-cppe5.md",{"path":326,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":327,"sourceHtmlUrl":328},"references\u002Ftao-rerun-segformer-foodseg103.md",3831,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ftao-rerun-segformer-foodseg103.md",{"path":330,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":331,"sourceHtmlUrl":332},"references\u002Ftao-rerun-smolvlm-vqav2.md",4534,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ftao-rerun-smolvlm-vqav2.md",{"path":334,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":335,"sourceHtmlUrl":336},"references\u002Ftesting.md",20681,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Ftesting.md",{"path":338,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":339,"sourceHtmlUrl":340},"references\u002Fvlm-scripts.md",27939,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Freferences\u002Fvlm-scripts.md",{"path":342,"resourceType":58,"resourceTypeDesc":59,"sizeBytes":343,"sourceHtmlUrl":344},"skill-card.md",3979,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fskill-card.md",{"path":346,"resourceType":68,"resourceTypeDesc":69,"sizeBytes":347,"sourceHtmlUrl":348},"skill.oms.sig",22217,"https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Ftao-finetune-huggingface-model\u002Fskill.oms.sig",[350,357,364,371,378,385],{"skillNo":351,"name":352,"nameZh":6,"description":353,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":354,"categories":355,"keywords":356,"publishedAt":19},"01M79G206MT3J6S0GY3C45P2KV","vss-summarize-video","Use to summarize a recorded video via the LVS summarization microservice (HITL-gated) with a VLM fallback. Not for report generation or live RTSP captioning.","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-summarize-video\u002FSKILL.md",[],[],{"skillNo":358,"name":359,"nameZh":6,"description":360,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":361,"categories":362,"keywords":363,"publishedAt":19},"01M79G1WN63PGRE8CKD47WEBQ5","vss-setup-video-analytics-api","Use to deploy the vss-video-analytics-api REST service standalone (config-source, data-log bind, Elasticsearch, optional Kafka). Not for full warehouse deploy.","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-setup-video-analytics-api\u002FSKILL.md",[],[],{"skillNo":365,"name":366,"nameZh":6,"description":367,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":368,"categories":369,"keywords":370,"publishedAt":19},"01M79G1S9F7VY3EF6H53PRP5AN","vss-setup-behavior-analytics","Use to deploy the vss-behavior-analytics service standalone (entrypoint, config-source, optional calibration). Not for the full warehouse deploy.","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-setup-behavior-analytics\u002FSKILL.md",[],[],{"skillNo":372,"name":373,"nameZh":6,"description":374,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":375,"categories":376,"keywords":377,"publishedAt":19},"01M79G1N1S1Y27PFZPYMNK1MTQ","vss-search-archive","Use this skill to run top-level VSS fusion search on archived video, or to ingest video files \u002F RTSP streams for search. Do NOT use for ad-hoc visual Q&A (use vss-ask-video), live captioning (use vss-deploy-dense-captioning), or video summarization and reports (use vss-summarize-video).","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-search-archive\u002FSKILL.md",[],[],{"skillNo":379,"name":380,"nameZh":6,"description":381,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":382,"categories":383,"keywords":384,"publishedAt":19},"01M79G1H6WN0Q6C0VCAXCZX6B7","vss-query-analytics","Use this skill when reading video-analytics metrics, incidents, alerts, and sensor data via the VA-MCP server (port 9901). Not for live VLM or incident-range narrative reports.","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-query-analytics\u002FSKILL.md",[],[],{"skillNo":386,"name":387,"nameZh":6,"description":388,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":389,"categories":390,"keywords":391,"publishedAt":19},"01M79G1DYPNXQS7QHZ0AAF5FX5","vss-manage-video-io-storage","Use to call the VIOS REST API (sensor list, timelines, clip extraction, snapshots, add\u002Fdelete sensors and streams). Not for VLM inference or search.","https:\u002F\u002Fgithub.com\u002FElogic360\u002Fsoftware-factory\u002Fblob\u002Fmain\u002Fskills\u002Fnvidia\u002Fvss-manage-video-io-storage\u002FSKILL.md",[],[]]