Workflow·Data Science & ML·v0.1.0

TAO Fine-tune HuggingFace Model

Name: TAO Fine-tune HuggingFace Model
Author: NVIDIA

Fine-tune any HuggingFace CV, VLM or LLM model on local NVIDIA GPUs inside an NGC PyTorch container.

You say

Buy it · $35 Read it before you buy $35 Written by NVIDIA · unverified publisher

Context cost: 99.6k tokensestimated from the bundle, loaded when it triggers
Bundle: 56 files · 398.3 kB17 scripts among them — read before you run
Licence: Apache-2.0paid listing
Last change: —v0.1.0
Servers it uses: Noneruns standalone

What it does

Fine-tune any HuggingFace CV / VLM / 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 / VLM / 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.

Installed, it changes the agent in these ways.

What this skill changes about the agent is not written down here yet. The listing was collected from its source, and the description is in its own SKILL.md.

Workflow

Runs a procedure end to end.

fine-tuningpytorchloratraining

Filed under

Data Science & ML

The skill itself

This is the whole product. A skill is instructions the model reads, so there is nothing behind the listing you cannot see first — the front matter loads with every session, and the body below it loads when the skill triggers.

SKILL.md18.0 kB · 373 lines

--- name: tao-finetune-huggingface-model description: > Fine-tune any HuggingFace CV / VLM / 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 / VLM / 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 / instance / panoptic segmentation, depth estimation, image-text-to-text VLM (SFT / LoRA), and LLM SFT / DPO / GRPO. Six-step workflow: inspect and qualify, hardware and NGC image, research, generate and smoke, train + eval + infer, push and emit rerun skill. license: Apache-2.0 tags: - finetuning - huggingface - nvidia-tao - computer-vision - training compatibility: 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/dataset is gated or `push_to_hub` is on; WANDB_API_KEY and WANDB_PROJECT only when WandB logging is enabled. metadata: author: NVIDIA Corporation version: "0.1.0" allowed-tools: Read Bash Write ---

30# tao-finetune-huggingface-model

32Local NVIDIA GPU fine-tuning for HuggingFace models, grounded in live-fetched

33documentation with curated references as a fallback safety net. One NGC container,

34a few focused scripts, one push to HF Hub. Follow the rules in this file; don't

35improvise.

37**Order of authority (highest first):**

391. **User input** — explicit model_id, dataset_id, training_method, config.yaml overrides.

402. **Live research** — model card, HF repo example, author finetune script, HF task docs, paper; always fetched (Step 3 + references/research-priorities.md).

413. **Curated references** (references/*.md) — fallback when live research is silent/ambiguous.

424. **Your training-data memory** — last resort; suspect, cross-check against (2)/(3).

44Conflict resolution between (2) and (3) and the source-line discrepancy note are

45in references/research-priorities.md.

47---

49## Inputs

51**Required:**

52- model_id — HuggingFace model ID, e.g. google/vit-base-patch16-224

54**Conditional credentials (read from the session environment, exported before launching when present):**

55- HF_TOKEN — only when the model/dataset 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" ].

56- WANDB_API_KEY, WANDB_PROJECT — only when WandB is enabled; WANDB_MODE=disabled opts out.

58**Dataset — exactly one:**

59- dataset_id — HuggingFace dataset ID *(source: hf)*

60- local_dataset_path — local folder or file *(source: local)*; optional

61 local_dataset_format ∈ {auto, imagefolder, coco, voc, jsonl, arrow, parquet,

62 csv} (default: auto-detect).

63- *(omit)* — agent recommends popular datasets *(source: recommend)*

65**Optional (have defaults):**

66- task_type — auto-detected from config + model card

67- n_train=10000, n_eval=1000, n_epochs=3, lora_r=16

68- output_dir=./output/<model_short_name>

69- hf_model_repo — push target; if unset and HF_TOKEN has write access,

70 auto-derived as <whoami>/<model_short_name>-finetuned.

71- push_to_hub=True — set to False to skip

72- skip_baseline=False — skip zero-shot baseline eval

74**Optional deliverables (off by default):**

75```yaml

76emit_progress_log: false # output_dir/PROGRESS.md (per-step journal)

77emit_report: false # reports/report.{pdf,html} with curves & samples

78emit_unit_tests: false # tests/ with fake-data heterogeneous-batch tests

79```

81All values live in output_dir/config.yaml. Never hardcode in Python.

83---

85## Execution platform

87This skill orchestrates *what* to run; the platform skills own *how* to run it on

88a GPU host — read them first.

90| Concern | Authoritative skill |

91|---|---|

92| GPU host runtime (driver 580, CUDA Toolkit 13.0, NVIDIA Container Toolkit 1.19.0) | [tao-skill-bank:tao-setup-nvidia-gpu-host](../../platform/tao-setup-nvidia-gpu-host/SKILL.md) |

93| docker run flags, NGC auth, mounts, env passthrough | [tao-skill-bank:tao-run-on-docker](../../platform/tao-run-on-docker/SKILL.md) |

94| Local Docker job preflight (daemon, GPU smoke) | [tao-skill-bank:tao-run-on-local-docker](../../platform/tao-run-on-local-docker/SKILL.md) |

96**Default platform:** local-docker — build a one-off image (run-<short>:latest)

97and run it on the local Docker daemon. Ask only when the user explicitly needs a

98different backend (Brev remote GPU, SLURM/Kubernetes); then run that platform's

99Preflight first and route the Steps 4–5 docker run commands through it. The

100GPU-runtime and presence-only credential preflights (values never read), the

101canonical docker run flag set, the list_tao_platforms.py selection command, and

102the workflow-specific flags (--entrypoint /bin/bash -lc, PYTORCH_CUDA_ALLOC_CONF,

103--name hft_train) are in references/workflow-intake-preflight.md.

104

105---

106

107## References — fallback safety net

108

109Consulted **only** when live research is silent, ambiguous, or unavailable; live

110docs always win for the specific model and current API. Each step links the

111references it needs; full catalog in references/detailed-workflow.md.

112

113Always-on: core-rules.md, error-playbook.md, compat-workarounds.md,

114model-discovery.md, dataset-recommendations.md, dataset-sources.md,

115dataset-patterns.md, hardware-container.md, research-priorities.md,

116cv-scripts.md, vlm-scripts.md, docker-runs.md, hub-push.md,

117pipeline-skill-template.md, deliverables.md. Opt-in (when their flag/need

118applies): progress-tracking.md, testing.md, reporting.md,

119workflow-intake-preflight.md, workflow-generate-train.md, workflow-push-rerun.md.

120

121**Rule:** before falling back, log the live source you tried and why it was

122insufficient (config.yaml notes:, and PROGRESS.md if enabled). [FETCH LIVE]

123markers in cv-scripts.md / vlm-scripts.md are a research checklist, not code to

124inline — refetch the listed URL if a block has no Step 3 finding.

125

126---

127

128## Core rules

129

130Non-negotiable behaviors. **Short version** (full enumeration —

131hallucinated-imports list, never-without-approval list, full error-recovery and

132hardware-sizing tables — in references/core-rules.md, consult before any

133training-time decision):

134

135- **Your HF-library knowledge is outdated.** Fetch live docs (model card, HF

136 repo example, task doc) before writing any ML code — don't generate trainer

137 args / collator / transforms from memory (Step 3).

138- **Smoke-test on real data with --max_steps 1** before any full run; no batch

139 launches without a verified smoke.

140- **Never silently substitute** model_id, dataset_id, or training_method — if

141 what the user asked for doesn't load, stop and ask.

142- **Error recovery is minimal-change.** OOM → halve batch, double grad_accum,

143 enable gradient checkpointing (no LoRA switch without approval); NaN → reduce

144 LR 10×; flat loss → inspect collator; same error 3× → stop and ask. Don't loop.

145- **Dataset columns verified BEFORE the collator** — rename in prepare_data.py;

146 restructuring needed → stop and ask.

147- **Hardware-sizing thumb (bf16):** ≤3B → 24 GB, 7–13B → 80 GB, 30B+ → multi-GPU

148 or LoRA on 1× 80 GB, 70B+ → 8× 80 GB or LoRA. Full finetune won't fit and no

149 LoRA requested → ask before switching.

150

151---

152

153## Workflow — 6 steps

154

155Single pass, sequential; each step has a clear gate before the next begins.

156

157### Step 1 — Inspect & qualify

158

159**Goal:** decide whether to proceed. Probe model + dataset, apply accept/reject,

160register applicable compat fixes, write the initial config.yaml.

161

162Prerequisites: MODEL_ID, optional DATASET_ID / local_dataset_path,

163optional HF_TOKEN, OUTPUT_DIR (default ./output/<model_short_name>). Probes

164run in a CPU-only python:3.12-slim Docker container (bind-mounted .probe/

165scratch) so the host needs no virtualenv — Docker must exist first. Docker-presence

166guard, container env, full probe invocation, and the model/dataset probe scripts

167are in references/workflow-intake-preflight.md, references/model-discovery.md,

168and references/dataset-sources.md.

169

170Probe requirements:

171

172- Model: load AutoConfig, read model-card tags, detect task from

173 architectures + tags + card examples (fallback logging in model-discovery.md).

174- Dataset: for recommended datasets, first present 3-5 choices from

175 dataset-recommendations.md; for local data, bind-mount read-only and use

176 dataset-sources.md format detection.

177- Reject early if the model config fails, the task is out of scope, no recipe

178 source exists, or the dataset cannot load / match the task schema.

179- Evaluate compat-workarounds.md against the model/task; defer hardware-dependent

180 rules to Step 2.

181

182Write the initial config.yaml (model_id, task, dataset_id or

183local_dataset_path, research_sources: [] filled in Step 3,

184applicable_workarounds: from Step 1, notes: [] for reference fallbacks,

185push_to_hub: true default — annotated template in

186references/workflow-intake-preflight.md). Optionally rm -rf "$OUTPUT_DIR/.probe"

187once the gate is met.

188

189**Gate:** config.yaml exists with model, dataset, task, applicable_workarounds;

190do not proceed if any field is missing.

191

192---

193

194### Step 2 — Hardware audit & NGC image

195

196**Goal:** verify Docker + GPU + disk, pick the NGC PyTorch image live, finalize

197hardware-dependent compat rules.

198

199**2a. Audit (hard gate)** — three checks (commands in

200references/workflow-intake-preflight.md):

2011. GPU host runtime — tao-setup-nvidia-gpu-host's

202 setup-nvidia-gpu-host.sh --backend docker --check-only; on fail, ask approval

203 then re-run with --install --yes.

2042. Free-disk soft-warn — override via MIN_DISK_GB (default 100 GB); recommend

205 ≥ 100 GB for NGC base (~20 GB) + HF cache + checkpoints + data.

2063. Conditional credential presence (from the session environment, values never

207 read) — HF_TOKEN only when gated or push_to_hub is on; WANDB_* only when

208 WandB is on.

209

210**Do not proceed to Step 4 on a hard-fail** — Step 4's docker build pulls a

21120+ GB NGC base, and a missing nvidia-container-toolkit only surfaces later as

212could not select device driver "" with capabilities: [[gpu]]. Record gpu_count,

213gpu_name, driver_major, vram_gb_per_gpu in config.yaml.

214

215**2b. Pick NGC image (live):** from the NVIDIA Deep Learning Frameworks support

216matrix (<https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html>),

217PyTorch NGC container section, pick the highest-versioned image where

218Min driver ≤ detected driver_major and container CUDA ≤ host CUDA Toolkit

219(match closely so cuDNN / TensorRT line up). Do **not** reject an image for an

220aN/bN/rcN PyTorch tag — NGC validates the full image; pick the newest

221CUDA-aligned one and let compat-workarounds.md handle per-version issues. If the

222matrix is unreachable, use the fallbacks in references/hardware-container.md;

223default nvcr.io/nvidia/pytorch:24.09-py3 (driver ≥ 545; SDPA+GQA bug — if

224num_key_value_heads < num_attention_heads, set attn_implementation: "eager").

225Record ngc_image in config.yaml.

226

227**2c. Re-evaluate hardware-dependent compat rules:** re-run the

228compat-workarounds.md walk for entries whose detect needs hw; update

229applicable_workarounds: in place.

230

231**2d. Model-fit check:** estimate param_bytes ≈ 2×param_count (bf16); if

232> 60% of vram_gb_per_gpu × 1e9, recommend LoRA in the user-facing summary.

233

234**Gate:** config.yaml has ngc_image, gpu_count, gpu_name, driver_major,

235vram_gb_per_gpu; hardware-dependent compat fixes recorded.

236

237---

238

239### Step 3 — Research the recipe

240

241**Goal:** fetch the live recipe — training-data knowledge of

242transformers/trl/peft is suspect, so Step 3 is non-negotiable. Walk

243references/research-priorities.md in priority order (Priority 1 → 6); stop once

244you have, for the detected task:

245

246- AutoModel / processor class

247- Train + eval transforms

248- Collator

249- compute_metrics

250- Hyperparameter hints (LR, batch size, epochs, scheduler)

251

252Record findings in meta/recipe.md, append source URLs to

253config.yaml: research_sources:. A slot with no live finding falls back to the

254matching scaffold (cv-scripts.md / vlm-scripts.md), logged as "fallback to

255scaffold — no live source for <slot>" under notes:. Conflict-resolution rules

256are in references/research-priorities.md.

257

258**Gate:** every required slot filled, with a source URL or scaffold-fallback note.

259

260---

261

262### Step 4 — Generate project & smoke-test

263

264**Goal:** write all scripts, build the image, prepare data, run a 1-step smoke on

265real data (one docker build, two docker runs).

266

267**4a. Generate project files** in output_dir/: config.yaml, Dockerfile,

268requirements.txt, prepare_data.py, train.py, run_eval.py, infer.py,

269optional merge_lora.py, optional tests/, .gitignore. Live Step 3 research is

270authority; cv-scripts.md / vlm-scripts.md give scaffold shape only. Apply every

271applicable_workarounds entry as a Dockerfile block, requirement pin, config

272override, or runtime env var. Hard rules: run_eval.py keeps that exact filename

273(avoids colliding with the HF evaluate package); every generated .py starts

274with the NVIDIA Apache-2.0 copyright header and any emitter fails when it is

275missing; emit_unit_tests: true generates and runs tests per

276references/testing.md. Script bodies, Dockerfile shape, and the emitter contract

277are in references/workflow-generate-train.md.

278

279**4b. Build, prepare, smoke** — docker build -t run-<short>:latest ., then

280prepare_data and the --smoke --max_steps 1 run (references/docker-runs.md

281§1-3). Smoke pass criteria (in logs/smoke.log):

282- No exception

283- Loss is finite (not 0.0, not NaN)

284- grad_norm > 0 at step 1

285

286If emit_unit_tests: true, also run pytest tests/ in the container. Any failure → STOP.

287

288**4c. Preflight summary** — before full training, print and verify: reference URL,

289dataset columns, Hub target, monitoring target, NGC image, hardware, smoke loss/grad norm.

290

291**Gate:** project files written, image built, smoke PASSED, preflight has no

292blank fields.

293

294---

295

296### Step 5 — Train, evaluate, infer

297

298**Goal:** baseline eval, full training, post-train eval, optional LoRA merge, 5

299inference samples (all commands: references/docker-runs.md §4-8).

300

301| Sub-step | docker-runs.md | Skip if |

302|---|---|---|

303| 5a. Baseline eval (zero-shot) | §4 | skip_baseline: true |

304| 5b. Full training (detached) | §5 | — |

305| 5c. LoRA merge | §6 | not VLM+LoRA |

306| 5d. Post-train eval | §7 | — |

307| 5e. Inference (5 samples) | §8 | — |

308

309Multi-GPU: prepend torchrun --nproc_per_node=$gpu_count to python train.py.

310

311While training streams, watch docker logs -f hft_train: loss should drop within

31210-20 steps; flat loss (collator/label-masking bug), NaN (LR too high), and OOM

313all stop the run — recovery in references/core-rules.md. If emit_report: true,

314run report.py after Step 5e per references/reporting.md.

315

316**Gate:** all of:

317- checkpoints/final/ (or checkpoints/merged/ for LoRA) exists

318- reports/eval_results.json has a numeric primary metric

319- reports/baseline_results.json exists (unless skipped)

320- reports/inference_samples/ has 5 samples

321- wandb URL shows descending loss

322

323---

324

325### Step 6 — Push & emit rerun skill

326

327**Goal:** publish the run and make it reproducible without re-research.

328

329Push per references/hub-push.md (weights, model card, eval/baseline JSONs,

330config.yaml, Dockerfile, requirements.txt, inference samples, reports when

331emitted) unless push_to_hub: false is explicit. Emit

332<output_dir>/skills/run-<short>/SKILL.md from

333references/pipeline-skill-template.md — substitute every placeholder, include

334full YAML metadata + the NVIDIA copyright HTML comment, and make any emitter fail

335if those are missing.

336

337**Gate (Done criteria):** all of:

338- Step 5 gate met

339- HF Hub repo exists at the resolved URL with weights + card + results/

340 (unless push_to_hub: false)

341- <output_dir>/skills/run-<short>/SKILL.md exists, no <placeholder> left,

342 with metadata + copyright HTML comment per pipeline-skill-template.md

343

344Final message: wandb URL, HF Hub URL, baseline -> fine-tuned primary metric,

345reports/inference_samples/, and the rerun skill path.

346

347---

348

349## Error playbook

350

351On a known runtime error, consult the symptom → minimal-fix table in

352references/error-playbook.md (NGC entrypoint, PyTorch/Transformers regressions,

353numpy ABI, Albumentations bbox, PEFT/checkpointing, LoRA target breadth, CV

354augmentation gaps, OOM at step 0) before redesigning anything. When a row there

355fires twice across runs, lift it into compat-workarounds.md with a detect rule

356— auto-applied in Step 1 before the error can fire.

357

358---

359

360## Communication style

361

362- Terse. No filler, no restating the request; one-word answers when appropriate.

363- Always include direct Hub and wandb URLs when referencing artifacts.

364- On error: state what went wrong, why, what you changed — no menus.

365- Never present "Option A/B/C" for a request with a clear answer. Act.

366

367## Example pipelines

368

369- [tao-rerun-convnext-cifar10](references/tao-rerun-convnext-cifar10.md)

370- [tao-rerun-detr-cppe5](references/tao-rerun-detr-cppe5.md)

371- [tao-rerun-segformer-foodseg103](references/tao-rerun-segformer-foodseg103.md)

372- [tao-rerun-smolvlm-vqav2](references/tao-rerun-smolvlm-vqav2.md)

373

In the file

SKILL.md2,306 words

Files56

LicenceApache-2.0

Why you can read it

Nothing in a skill executes. The client loads the text and the model follows it, so a skill can be audited the way a runbook is — by reading it.

What it costs in context

Skills are not billed by the call. They are paid for in context: every token the instructions occupy is a token your code, your diff and your conversation cannot use. Here is what this one takes and when it takes it.

≈350

always loaded

The name and description, so the model knows the skill exists and when to reach for it.

99,225

on trigger

The instruction body and 55 supporting files, read only when the skill fires.

49.8%

of a 200k window

Ten skills this size would take about 498% of the window before you open a file.

050k100k150k200k context window

99.6k tokens, estimated from the bundle at four bytes to the token, held for the rest of the session once it triggers. Heavy. Teams tend to install this one per project rather than globally, and load it only when the job comes up.

Servers bill, skills cost

A server charges by the month. A skill charges once per session, in context, and then keeps charging it for as long as the session lives.

Before and after

The same question, put to the same model twice: once as it comes, and once with these instructions loaded.

No worked example has been published for this skill yet.

Adoption

Installsnone yet

Ratingno reviews yet

The procedure it runs

The procedure has not been published here. It is in the skill’s own SKILL.md, which its author has not sent to the marketplace yet.

Prose, not code

These steps are written for a model to follow, not executed by a runtime. It can still be told to skip one, and it will say so when it does.

Servers it uses

None. This skill calls no MCP servers at all.

Everything it needs is in the instructions, so it works in a project with nothing connected — the model reads the file and changes how it works with what it can already reach.

It asks the agent to write files, using whatever file access your client already has. It never touches the network.

What it asks for

Writes filesyes

Network accessno

Read from the allowed-tools line of this skill’s own SKILL.md. A skill grants no permissions of its own — it can only ask for tools your client already has.

What it will not do

Every skill is narrow, and the useful ones say where they stop. These are the jobs this one is the wrong tool for.

What this skill is not for has not been published here. Nothing is implied by that: it is a section the author has not filled in.

If this is not it

12 other skills in Data Science & ML

What is in the bundle

56 files, 398.3 kB on disk. Mostly text — the instructions the model reads — with 17 scripts in it that your client would run only if the instructions tell it to.

BENCHMARK.md4.2 kB
SKILL.md18.0 kB
skill-card.md4.2 kB
evals/evals.json0.8 kB
examples/README.md6.3 kB
examples/convnext-tiny-cifar10/config.yaml2.1 kB
examples/convnext-tiny-cifar10/infer.py2.5 kB
examples/convnext-tiny-cifar10/prepare_data.py1.4 kB
examples/convnext-tiny-cifar10/requirements.txt0.9 kB
examples/convnext-tiny-cifar10/run_eval.py3.1 kB
examples/convnext-tiny-cifar10/train.py5.6 kB
examples/detr-resnet50-cppe5/config.yaml2.1 kB
examples/detr-resnet50-cppe5/infer.py3.5 kB
examples/detr-resnet50-cppe5/prepare_data.py1.3 kB
examples/detr-resnet50-cppe5/requirements.txt1.0 kB
examples/detr-resnet50-cppe5/run_eval.py3.7 kB
examples/detr-resnet50-cppe5/train.py9.2 kB
examples/segformer-b0-foodseg103/config.yaml2.2 kB
examples/segformer-b0-foodseg103/infer.py3.3 kB
examples/segformer-b0-foodseg103/prepare_data.py1.1 kB
examples/segformer-b0-foodseg103/requirements.txt0.7 kB
examples/segformer-b0-foodseg103/run_eval.py3.1 kB
examples/segformer-b0-foodseg103/train.py6.7 kB
examples/smolvlm-256m-vqav2/config.yaml2.2 kB
examples/smolvlm-256m-vqav2/infer.py2.9 kB
examples/smolvlm-256m-vqav2/merge_lora.py1.2 kB
examples/smolvlm-256m-vqav2/prepare_data.py1.2 kB
examples/smolvlm-256m-vqav2/requirements.txt0.9 kB
examples/smolvlm-256m-vqav2/run_eval.py3.0 kB
examples/smolvlm-256m-vqav2/train.py5.0 kB
references/compat-workarounds.md12.1 kB
references/core-rules.md4.8 kB
references/cv-scripts.md34.0 kB
references/dataset-patterns.md10.8 kB
references/dataset-recommendations.md11.3 kB
references/dataset-sources.md17.7 kB
references/deliverables.md20.6 kB
references/detailed-workflow.md0.7 kB
references/docker-runs.md9.1 kB
references/error-playbook.md3.8 kB
references/hardware-container.md8.7 kB
references/hub-push.md5.6 kB
references/model-discovery.md10.8 kB
references/pipeline-skill-template.md9.2 kB
references/progress-tracking.md9.9 kB
references/reporting.md18.9 kB
references/research-priorities.md6.6 kB
references/tao-rerun-convnext-cifar10.md4.6 kB
references/tao-rerun-detr-cppe5.md3.9 kB
references/tao-rerun-segformer-foodseg103.md3.8 kB
references/tao-rerun-smolvlm-vqav2.md4.5 kB
references/testing.md21.3 kB
references/vlm-scripts.md28.4 kB
references/workflow-generate-train.md7.7 kB
references/workflow-intake-preflight.md23.0 kB
references/workflow-push-rerun.md3.1 kB

What is not in it

A skill installs nothing and depends on nothing: it is a folder your client reads. This one carries 17 scripts beside the text, so the bundle is 56 files you can review in full before installing. The Apache-2.0 licence covers the templates and examples as well as the instructions.

Install

Installing copies the bundle into your project. Nothing runs at install time — the files sit on disk until the model reads them.

$35 once

TAO Fine-tune HuggingFace Model · Apache-2.0 · NVIDIA

one-time

Price$35 once

LicenceApache-2.0 — the author’s, unchanged by this purchase

Paid throughStripe, once, on the card you add at the checkout

Keeps workingfor good — the files are yours once they are on disk

Updatesevery release of 0.x through this account

You can read the whole bundle before paying — the SKILL.md above is the product, not a preview of it. What the money buys is the delivery: the folder packaged and handed to your machine by key, every update its author ships, and our support if it does not do what this listing says. The terms of use are Apache-2.0, set by the author and unchanged by buying it here.

Payment runs through Stripe, on a page like this one rather than a redirect. Once there is an account it joins the same mcprush invoice as everything else you run, so there is never a second card to enter.

Which clients pick it up on their own

A skill is a folder of text. A client with a skills folder reads it without being told; everywhere else the same text works, it is just handed to the model rather than found.

Claude Code.claude/skills/

Claude Desktop

ChatGPT

Cursor.cursor/skills/

VS Code.github/skills/

Codex CLI.agents/skills/

Gemini CLI.gemini/skills/

Grok.grok/skills/

Zed.agents/skills/

Windsurf.windsurf/skills/

Agent SDK.claude/skills/

HTTP / API

This release

Version0.1.0

Publishedno release date on file

Price$35

Referencenvidia/tao-fine-tune-huggingface-model

Versions

v0.1.0 is what is on the shelf; no release here carries a date. Instructions change more often than APIs do — a skill can be rewritten entirely without anything it depends on moving.

v0.1.0

No earlier releases have been published to the marketplace.

Pinning

Put nvidia/tao-fine-tune-huggingface-model@0.1.0 in the install command to hold this exact version. Without the suffix you get whatever is current the day you install, and nothing moves under you afterwards.

Reviews

—no reviews yet · no installs yet

Nobody has reviewed this skill yet. The rating is the mean of the reviews written here, so there is none until somebody writes the first.

Who can post

Only accounts that have had the skill installed for fourteen days, so a review is written after living with it rather than after reading it. Publishers may reply once.

Who wrote it

NVIDIA

Publishes on mcprush.

0 servers listed9 skills listednot claimed

Profile

Publisher

Servers0