Vision Model Training on Hugging Face Jobs workflow skill. Use this skill when the user needs Train or fine-tune vision models on Hugging Face Jobs for detection, classification, and SAM or SAM2 segmentation and the operator should preserve the upstream workflow, copied support files, and provenance before merging or handing off.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Vision Model Training on Hugging Face Jobs workflow skill. Use this skill when the user needs Train or fine-tune vision models on Hugging Face Jobs for detection, classification, and SAM or SAM2 segmentation and the operator should preserve the upstream workflow, copied support files, and provenance before merging or handing off.
This public intake copy packages plugins/antigravity-awesome-skills/skills/hugging-face-vision-trainer from https://github.com/sickn33/antigravity-awesome-skills into the native Omni Skills editorial shape without hiding its origin.
Use it when the operator needs the upstream workflow, support files, and repository context to stay intact while the public validator and private enhancer continue their normal downstream flow.
This intake keeps the copied upstream files intact and uses the external_source block in metadata.json plus ORIGIN.md as the provenance anchor for review.
Vision Model Training on Hugging Face Jobs Train object detection, image classification, and SAM/SAM2 segmentation models on managed cloud GPUs. No local GPU setup required—results are automatically saved to the Hugging Face Hub.
Imported source sections that did not map cleanly to the public headings are still preserved below or in the support files. Notable imported sections: Local Script Execution, Prerequisites Checklist, Dataset Validation, Critical directives, Model & hardware selection, Quick start — Object Detection.
When to Use This Skill
Use this section as the trigger filter. It should make the activation boundary explicit before the operator loads files, runs commands, or opens a pull request.
Fine-tune object detection models (D-FINE, RT-DETR v2, DETR, YOLOS) on cloud GPUs or local
Fine-tune image classification models (timm: MobileNetV3, MobileViT, ResNet, ViT/DINOv3, or any Transformers classifier) on cloud GPUs or local
Fine-tune SAM or SAM2 models for segmentation / image matting using bbox or point prompts
Train bounding-box detectors on custom datasets
Train image classifiers on custom datasets
Train segmentation models on custom mask datasets with prompts
Operating Table
Situation
Start here
Why it matters
First-time use
metadata.json
Confirms repository, branch, commit, and imported path through the external_source block before touching the copied workflow
Provenance review
ORIGIN.md
Gives reviewers a plain-language audit trail for the imported source
Workflow execution
references/finetune_sam2_trainer.md
Starts with the smallest copied file that materially changes execution
Supporting context
references/hub_saving.md
Adds the next most relevant copied source file without loading the entire package
Handoff decision
## Related Skills
Helps the operator switch to a stronger native skill when the task drifts
Workflow
This workflow is intentionally editorial and operational at the same time. It keeps the imported source useful to the operator while still satisfying the public intake standards that feed the downstream enhancer flow.
Step 5: Save script locally, submit job, and report details
Save the script locally to submittedjobs/ in the workspace root (create if needed) with a descriptive name like training.py. Tell the user the path.
Submit using hfjobs MCP tool (preferred) or HfApi().runuvjob() — see directive #1 for both methods. Pass all config via scriptargs.
Imported Workflow Notes
Imported: Automatic Bbox Preprocessing
The object detection training script (scripts/object_detection_training.py) automatically handles bbox format detection (xyxy→xywh conversion), bbox sanitization, image_id generation, string category→integer remapping, and dataset truncation. No manual preprocessing needed — just ensure the dataset has objects.bbox and objects.category columns.
Imported: Training workflow
Copy this checklist and track progress:
Training Progress:
- [ ] Step 1: Verify prerequisites (account, token, dataset)
- [ ] Step 2: Validate dataset format (run dataset_inspector.py)
- [ ] Step 3: Ask user about dataset size and validation split
- [ ] Step 4: Prepare training script (OD: scripts/object_detection_training.py, IC: scripts/image_classification_training.py, SAM: scripts/sam_segmentation_training.py)
- [ ] Step 5: Save script locally, submit job, and report details
Step 1: Verify prerequisites
Follow the Prerequisites Checklist above.
Step 2: Validate dataset
Run the dataset inspector BEFORE spending GPU time. See "Dataset Validation" section above.
Step 3: Ask user preferences
ALWAYS use the AskUserQuestion tool with option-style format:
AskUserQuestion({
"questions": [
{
"question": "Do you want to run a quick test with a subset of the data first?",
"header": "Dataset Size",
"options": [
{"label": "Quick test run (10% of data)", "description": "Faster, cheaper (~30-60 min, ~$2-5) to validate setup"},
{"label": "Full dataset (Recommended)", "description": "Complete training for best model quality"}
],
"multiSelect": false
},
{
"question": "Do you want to create a validation split from the training data?",
"header": "Split data",
"options": [
{"label": "Yes (Recommended)", "description": "Automatically split 15% of training data for validation"},
{"label": "No", "description": "Use existing validation split from dataset"}
],
"multiSelect": false
},
{
"question": "Which GPU hardware do you want to use?",
"header": "Hardware Flavor",
"options": [
{"label": "t4-small ($0.40/hr)", "description": "1x T4, 16 GB VRAM — sufficient for all OD models under 100M params"},
{"label": "l4x1 ($0.80/hr)", "description": "1x L4, 24 GB VRAM — more headroom for large images or batch sizes"},
{"label": "a10g-large ($1.50/hr)", "description": "1x A10G, 24 GB VRAM — faster training, more CPU/RAM"},
{"label": "a100-large ($2.50/hr)", "description": "1x A100, 80 GB VRAM — fastest, for very large datasets or image sizes"}
],
"multiSelect": false
}
]
})
Save the script locally to submitted_jobs/ in the workspace root (create if needed) with a descriptive name like training_<dataset>_<YYYYMMDD_HHMMSS>.py. Tell the user the path.
Submit using hf_jobs MCP tool (preferred) or HfApi().run_uv_job() — see directive #1 for both methods. Pass all config via script_args.
Report the job ID (from .id attribute), monitoring URL, Trackio dashboard (https://huggingface.co/spaces/{username}/trackio), expected time, and estimated cost.
Wait for user to request status checks — don't poll automatically. Training jobs run asynchronously and can take hours.
Imported: Local Script Execution
Helper scripts use PEP 723 inline dependencies. Run them with uv run:
uv run scripts/dataset_inspector.py --dataset username/dataset-name --split train
uv run scripts/estimate_cost.py --help
Examples
Example 1: Ask for the upstream workflow directly
Use @hugging-face-vision-trainer-v2 to handle <task>. Start from the copied upstream workflow, load only the files that change the outcome, and keep provenance visible in the answer.
Explanation: This is the safest starting point when the operator needs the imported workflow, but not the entire repository.
Example 2: Ask for a provenance-grounded review
Review @hugging-face-vision-trainer-v2 against metadata.json and ORIGIN.md, then explain which copied upstream files you would load first and why.
Explanation: Use this before review or troubleshooting when you need a precise, auditable explanation of origin and file selection.
Example 3: Narrow the copied support files before execution
Use @hugging-face-vision-trainer-v2 for <task>. Load only the copied references, examples, or scripts that change the outcome, and name the files explicitly before proceeding.
Explanation: This keeps the skill aligned with progressive disclosure instead of loading the whole copied package by default.
Example 4: Build a reviewer packet
Review @hugging-face-vision-trainer-v2 using the copied upstream files plus provenance, then summarize any gaps before merge.
Explanation: This is useful when the PR is waiting for human review and you want a repeatable audit packet.
Best Practices
Treat the generated public skill as a reviewable packaging layer around the upstream repository. The goal is to keep provenance explicit and load only the copied source material that materially improves execution.
Keep the imported skill grounded in the upstream repository; do not invent steps that the source material cannot support.
Prefer the smallest useful set of support files so the workflow stays auditable and fast to review.
Keep provenance, source commit, and imported file paths visible in notes and PR descriptions.
Point directly at the copied upstream files that justify the workflow instead of relying on generic review boilerplate.
Treat generated examples as scaffolding; adapt them to the concrete task before execution.
Route to a stronger native skill when architecture, debugging, design, or security concerns become dominant.
Troubleshooting
Problem: The operator skipped the imported context and answered too generically
Symptoms: The result ignores the upstream workflow in plugins/antigravity-awesome-skills/skills/hugging-face-vision-trainer, fails to mention provenance, or does not use any copied source files at all.
Solution: Re-open metadata.json, ORIGIN.md, and the most relevant copied upstream files. Check the external_source block first, then restate the provenance before continuing.
Problem: The imported workflow feels incomplete during review
Symptoms: Reviewers can see the generated SKILL.md, but they cannot quickly tell which references, examples, or scripts matter for the current task.
Solution: Point at the exact copied references, examples, scripts, or assets that justify the path you took. If the gap is still real, record it in the PR instead of hiding it.
Problem: The task drifted into a different specialization
Symptoms: The imported skill starts in the right place, but the work turns into debugging, architecture, design, security, or release orchestration that a native skill handles better.
Solution: Use the related skills section to hand off deliberately. Keep the imported provenance visible so the next skill inherits the right context instead of starting blind.
Imported Troubleshooting Notes
Imported: Common failure modes
OOM (CUDA out of memory)
Reduce per_device_train_batch_size (try 4, then 2), reduce IMAGE_SIZE, or upgrade hardware.
Dataset format errors
Run scripts/dataset_inspector.py first. The training script auto-detects xyxy vs xywh, converts string categories to integer IDs, and adds image_id if missing. Ensure objects.bbox contains 4-value coordinate lists in absolute pixels and objects.category contains either integer IDs or string labels.
Hub push failures (401)
Verify: (1) job secrets include token (see directive #2), (2) script sets training_args.hub_token BEFORE creating the Trainer, (3) push_to_hub=True is set, (4) correct hub_model_id, (5) token has write permissions.
Job timeout
Increase timeout (see directive #5 table), reduce epochs/dataset, or use checkpoint strategy with hub_strategy="every_save".
KeyError: 'test' (missing test split)
The object detection training script handles this gracefully — it falls back to the validation split. Ensure you're using the latest scripts/object_detection_training.py.
Single-class dataset: "iteration over a 0-d tensor"
torchmetrics.MeanAveragePrecision returns scalar (0-d) tensors for per-class metrics when there's only one class. The template scripts/object_detection_training.py handles this by calling .unsqueeze(0) on these tensors. Ensure you're using the latest template.
Poor detection performance (mAP < 0.15)
Increase epochs (30-50), ensure 500+ images, check per-class mAP for imbalanced classes, try different learning rates (1e-5 to 1e-4), increase image size.
@00-andruia-consultant - Use when the work is better handled by that native specialization after this imported skill establishes context.
@00-andruia-consultant-v2 - Use when the work is better handled by that native specialization after this imported skill establishes context.
@10-andruia-skill-smith - Use when the work is better handled by that native specialization after this imported skill establishes context.
@10-andruia-skill-smith-v2 - Use when the work is better handled by that native specialization after this imported skill establishes context.
Additional Resources
Use this support matrix and the linked files below as the operator packet for this imported skill. They should reflect real copied source material, not generic scaffolding.
Resource family
What it gives the reviewer
Example path
references
copied reference notes, guides, or background material from upstream
references/finetune_sam2_trainer.md
examples
worked examples or reusable prompts copied from upstream
examples/n/a
scripts
upstream helper scripts that change execution or validation
scripts/dataset_inspector.py
agents
routing or delegation notes that are genuinely part of the imported package
agents/n/a
assets
supporting assets or schemas copied from the source package
If hf_jobs MCP tool is unavailable, use the Python API directly:
from huggingface_hub import HfApi, get_token
api = HfApi()
job_info = api.run_uv_job(
script="path/to/training_script.py", # file PATH, NOT content
script_args=["--dataset_name", "cppe-5", ...],
flavor="a10g-large",
timeout=14400, # seconds (4 hours)
env={"PYTHONUNBUFFERED": "1"},
secrets={"HF_TOKEN": get_token()}, # MUST use get_token(), NOT "$HF_TOKEN"
)
print(f"Job ID: {job_info.id}")
Critical differences between the two methods:
hf_jobs MCP tool
HfApi().run_uv_job()
script param
Python code string or URL (NOT local paths)
File path to .py file (NOT content)
Token in secrets
"$HF_TOKEN" (auto-replaced)
get_token() (actual token value)
Timeout format
String ("4h")
Seconds (14400)
Rules for both methods:
The training script MUST include PEP 723 inline metadata with dependencies
Do NOT use image or command parameters (those belong to run_job(), not run_uv_job())
2. Authentication via job secrets + explicit hub_token injection
Job config MUST include the token in secrets — syntax depends on submission method (see table above).
Training script requirement: The Transformers Trainer calls create_repo(token=self.args.hub_token) during __init__() when push_to_hub=True. The training script MUST inject HF_TOKEN into training_args.hub_token AFTER parsing args but BEFORE creating the Trainer. The template scripts/object_detection_training.py already includes this:
hf_token = os.environ.get("HF_TOKEN")
if training_args.push_to_hub andnot training_args.hub_token:
if hf_token:
training_args.hub_token = hf_token
If you write a custom script, you MUST include this token injection before the Trainer(...) call.
Do NOT call login() in custom scripts unless replicating the full pattern from scripts/object_detection_training.py
Do NOT rely on implicit token resolution (hub_token=None) — unreliable in Jobs
See the hugging-face-jobs skill → Token Usage Guide for full details
3. JobInfo attribute
Access the job identifier using .id (NOT .job_id or .name — these don't exist):
job_info = api.run_uv_job(...) # or hf_jobs("uv", {...})
job_id = job_info.id# Correct -- returns string like "687fb701029421ae5549d998"
4. Required training flags and HfArgumentParser boolean syntax
scripts/object_detection_training.py uses HfArgumentParser — all config is passed via script_args. Boolean arguments have two syntaxes:
bool fields (e.g., push_to_hub, do_train): Use as bare flags (--push_to_hub) or negate with --no_ prefix (--no_remove_unused_columns)
Optional[bool] fields (e.g., greater_is_better): MUST pass explicit value (--greater_is_better True). Bare --greater_is_better causes error: expected one argument
Required flags for object detection:
--no_remove_unused_columns # MUST: preserves image column for pixel_values
--no_eval_do_concat_batches # MUST: images have different numbers of target boxes
--push_to_hub # MUST: environment is ephemeral
--hub_model_id username/model-name
--metric_for_best_model eval_map
--greater_is_better True # MUST pass "True" explicitly (Optional[bool])
--do_train
--do_eval
Required flags for image classification:
--no_remove_unused_columns # MUST: preserves image column for pixel_values
--push_to_hub # MUST: environment is ephemeral
--hub_model_id username/model-name
--metric_for_best_model eval_accuracy
--greater_is_better True # MUST pass "True" explicitly (Optional[bool])
--do_train
--do_eval
Required flags for SAM/SAM2 segmentation:
--remove_unused_columns False # MUST: preserves input_boxes/input_points
--push_to_hub # MUST: environment is ephemeral
--hub_model_id username/model-name
--do_train
--prompt_type bbox # or "point"
--dataloader_pin_memory False # MUST: avoids pin_memory issues with custom collator
5. Timeout management
Default 30 min is TOO SHORT for object detection. Set minimum 2-4 hours. Add 30% buffer for model loading, preprocessing, and Hub push.
Scenario
Timeout
Quick test (100-200 images, 5-10 epochs)
1h
Development (500-1K images, 15-20 epochs)
2-3h
Production (1K-5K images, 30 epochs)
4-6h
Large dataset (5K+ images)
6-12h
6. Trackio monitoring
Trackio is always enabled in the object detection training script — it calls trackio.init() and trackio.finish() automatically. No need to pass --report_to trackio. The project name is taken from --output_dir and the run name from --run_name. For image classification, pass --report_to trackio in TrainingArguments.
Start with ustc-community/dfine-small-coco for fast iteration. Move to D-FINE Large or RT-DETR v2 R50 for better accuracy.
Recommended image classification models
All timm/ models work out of the box via AutoModelForImageClassification (loaded as TimmWrapperForImageClassification). See references/timm_trainer.md for details.
Model
Params
Use case
timm/mobilenetv3_small_100.lamb_in1k
2.5M
Ultra-lightweight — mobile/edge, fastest training
timm/mobilevit_s.cvnets_in1k
5.6M
Mobile transformer — good accuracy/speed trade-off
timm/resnet50.a1_in1k
25.6M
Strong CNN baseline — reliable, well-studied
timm/vit_base_patch16_dinov3.lvd1689m
86.6M
Best accuracy — DINOv3 self-supervised ViT
Start with timm/mobilenetv3_small_100.lamb_in1k for fast iteration. Move to timm/resnet50.a1_in1k or timm/vit_base_patch16_dinov3.lvd1689m for better accuracy.
Recommended SAM/SAM2 segmentation models
Model
Params
Use case
facebook/sam2.1-hiera-tiny
38.9M
Fastest SAM2 — good for quick experiments
facebook/sam2.1-hiera-small
46.0M
Best starting point — good quality/speed balance
facebook/sam2.1-hiera-base-plus
80.8M
Higher capacity for complex segmentation
facebook/sam2.1-hiera-large
224.4M
Best SAM2 accuracy — requires more VRAM
facebook/sam-vit-base
93.7M
Original SAM — ViT-B backbone
facebook/sam-vit-large
312.3M
Original SAM — ViT-L backbone
facebook/sam-vit-huge
641.1M
Original SAM — ViT-H, best SAM v1 accuracy
Start with facebook/sam2.1-hiera-small for fast iteration. SAM2 models are generally more efficient than SAM v1 at similar quality. Only the mask decoder is trained by default (vision and prompt encoders are frozen).
Hardware recommendation
All recommended OD and IC models are under 100M params — t4-small (16 GB VRAM, $0.40/hr) is sufficient for all of them. Image classification models are generally smaller and faster than object detection models — t4-small handles even ViT-Base comfortably. For SAM2 models up to hiera-base-plus, t4-small is sufficient since only the mask decoder is trained. For sam2.1-hiera-large or SAM v1 models, use l4x1 or a10g-large. Only upgrade if you hit OOM from large batch sizes — reduce batch size first before switching hardware. Common upgrade path: t4-small → l4x1 ($0.80/hr, 24 GB) → a10g-large ($1.50/hr, 24 GB).
For full hardware flavor list: refer to the hugging-face-jobs skill. For cost estimation: run scripts/estimate_cost.py.
Imported: Quick start — Object Detection
The script_args below are the same for both submission methods. See directive #1 for the critical differences between them.
--model_name_or_path — SAM or SAM2 model (see model table above); auto-detects SAM vs SAM2
--dataset_name — the Hub dataset ID (e.g., "merve/MicroMat-mini")
--prompt_type — "bbox" or "point" — type of prompt in the dataset
--prompt_column_name — column with JSON-encoded prompts (default: "prompt")
--bbox_column_name — dedicated bbox column (alternative to JSON prompt column)
--point_column_name — dedicated point column (alternative to JSON prompt column)
--mask_column_name — column with ground-truth masks (default: "mask")
--hub_model_id — "username/model-name" for Hub persistence
--num_train_epochs — 20-30 typical for SAM fine-tuning
--per_device_train_batch_size — 2-4 (SAM models use significant memory)
--freeze_vision_encoder / --freeze_prompt_encoder — freeze encoder weights (default: both frozen, only mask decoder trains)
--train_val_split — fraction to split for validation (default 0.1)
Imported: Checking job status
MCP tool (if available):
hf_jobs("ps") # List all jobs
hf_jobs("logs", {"job_id": "your-job-id"}) # View logs
hf_jobs("inspect", {"job_id": "your-job-id"}) # Job details
Python API fallback:
from huggingface_hub import HfApi
api = HfApi()
api.list_jobs() # List all jobs
api.get_job_logs(job_id="your-job-id") # View logs
api.get_job(job_id="your-job-id") # Job details
