무료GitHub에 공개

unsloth-hf-jobs

Fine-tune LLMs and VLMs using Unsloth on HF Jobs (Hugging Face on-demand cloud GPUs). Use when users want to fine-tune language models, train VLMs (Vision Language Models), do continued pretraining, domain adaptation, or run UV scripts on HF Jobs. Triggers on requests involving Unsloth training, HF Jobs GPU training, Qwen3-VL fine-tuning, Gemma VLM training, or LoRA fine-tuning on cloud GPUs.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills

연동 도구Hugging Face

실행 점수 미검증

설치 방법 보기

포함된 파일(9)

SKILL.md5.2 KB
eval.yaml986 B
graders/check.py3.9 KB
instructions/qwen3-vl-finetune-hf-jobs.md1.5 KB
rubrics/quality.md1.5 KB
scripts/continued-pretraining.py12.0 KB
scripts/sft-gemma3-vlm.py11.0 KB
scripts/sft-qwen3-vl.py19.6 KB
solutions/reference-qwen3-vl-finetune-hf-jobs1.6 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Unsloth Training on HF Jobs

Fine-tune LLMs and VLMs using Unsloth on HF Jobs with UV scripts that handle dependencies automatically.

Prerequisites

Hugging Face account with token
HF CLI: curl -LsSf https://hf.co/cli/install.sh | bash
Dataset on the Hub (see data formats below)

Available Scripts

Script	Base Model	Task
`scripts/sft-qwen3-vl.py`	Qwen3-VL-8B	VLM fine-tuning
`scripts/sft-gemma3-vlm.py`	Gemma 3 4B	VLM fine-tuning (smaller)
`scripts/continued-pretraining.py`	Qwen3-0.6B	Domain adaptation

Remote URL (for HF Jobs): https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/

Data Formats

VLM Fine-tuning

Requires images and messages columns:

{
    "images": [<PIL.Image>],
    "messages": [
        {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What's in this image?"}]},
        {"role": "assistant", "content": [{"type": "text", "text": "A golden retriever playing fetch."}]}
    ]
}

Example dataset: davanstrien/iconclass-vlm-sft

Continued Pretraining

Any dataset with a text column (use --text-column if named differently):

{"text": "Your domain-specific text here..."}

Usage Patterns

VLM Fine-tuning

hf jobs uv run \
  https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/sft-qwen3-vl.py \
  --flavor a100-large --secrets HF_TOKEN --timeout 4h \
  -- --dataset <username>/<dataset> \
     --num-epochs 1 \
     --eval-split 0.2 \
     --output-repo <username>/<model-name>

Continued Pretraining

hf jobs uv run \
  https://huggingface.co/datasets/uv-scripts/unsloth-jobs/raw/main/continued-pretraining.py \
  --flavor a100-large --secrets HF_TOKEN \
  -- --dataset <username>/<dataset> \
     --text-column content \
     --max-steps 1000 \
     --output-repo <username>/<model-name>

With Trackio Monitoring

Add --trackio-space <username>/trackio to any command for live monitoring.

Common Options

Option	Description	Default
`--dataset`	HF dataset ID	required
`--output-repo`	Where to save trained model	required
`--max-steps`	Training steps	500
`--num-epochs`	Train for N epochs (overrides steps)	-
`--eval-split`	Fraction for evaluation	0
`--batch-size`	Per-device batch size	2
`--gradient-accumulation`	Accumulation steps	4
`--lora-r`	LoRA rank	16
`--learning-rate`	Learning rate	2e-4
`--merge-model`	Upload merged model (not just adapter)	false
`--streaming`	Stream large datasets	false

Quick Tips

Verify setup with --max-steps 10 before full runs
Use --eval-split 0.1 to detect overfitting
Check GPU pricing: hf jobs hardware (A100-large ~$2.50/hr, L40S ~$1.80/hr)
First step may take minutes (CUDA kernel compilation)
View script options: uv run <script-url> --help

Anti-Patterns to Avoid

Skipping the smoke test — always verify setup with --max-steps 10 before a full run; config errors cost GPU hours
Wrong data format for VLMs — VLM fine-tuning requires images + messages columns with image + text content parts; plain text datasets silently fail or train nonsense
Forgetting --secrets HF_TOKEN — the job fails at model download/login; pass the secret explicitly
Batch size 1 without gradient accumulation — on A100 with 4-bit LoRA you can afford larger batches; accumulation of 1 wastes throughput
Uploading only the adapter when you need a merged model — set --merge-model if the target is deployment, not another fine-tuning pass
Training without --eval-split — you cannot detect overfitting or pick a checkpoint without a held-out split
Ignoring --streaming on huge datasets — disk may fill up; stream when the dataset doesn't fit locally

When to Use / Not Use

Use this skill when:

Fine-tuning LLMs or VLMs with Unsloth on HF on-demand cloud GPUs
Doing LoRA/QLoRA fine-tuning, continued pretraining, or domain adaptation via UV scripts
You have no local GPU or want ephemeral, pay-per-hour training

Do NOT use this skill when:

You have local GPUs and prefer on-premises training — use the unsloth skill directly
The model works with prompting/RAG and doesn't need fine-tuning at all
You need full control over a custom training stack (TRL, custom loops) — this skill is scoped to the provided UV scripts

Resources

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

같은 저장소의 스킬

adaption-ai

무료

Adaption AI SDK for synthetic data augmentation and dataset adaptation. Use when building data pipelines with the Adaption Python SDK, uploading datasets (local files, Hugging Face, Kaggle), running augmentation/adaptation jobs, configuring brand controls (hallucination mitigation, safety categories, length), recipe specifications (reasoning traces, deduplication, preference pairs, prompt rephrase), evaluating dataset quality, downloading results, or any workflow involving `pip install adaption`, `from adaption import Adaption`, Adaptive Data, or the adaptionlabs.ai API. Also trigger when the user mentions synthetic data generation for fine-tuning, dataset augmentation pipelines, DPO preference pair generation, or grounding-based hallucination reduction on training data.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

api-design

무료

Design and review intuitive, scalable, maintainable HTTP APIs. Use this skill whenever the user wants to design a new REST API, review an existing API or spec, write OpenAPI 3.x definitions, or work with HTTP semantics (GET/POST/PUT/PATCH/DELETE), status codes, idempotency (Idempotency-Key), error envelopes (RFC 7807 problem+json), pagination, filtering, versioning, or API auth (API keys, OAuth2 client credentials, rate limits). Also trigger for "API design", "RESTful", "endpoints", "OpenAPI", "Swagger", "ReDoc", "contract testing", or GraphQL and gRPC design questions.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

database-design

무료

Design relational database schemas (and choose when to go NoSQL) that stay maintainable and fast. Use this skill whenever the user mentions tables, DDL, entities and relationships, normalization (1NF/2NF/3NF), primary and foreign keys, UUID vs bigint IDs, indexes (B-tree, composite, covering, partial), EXPLAIN, constraints (CHECK, UNIQUE, exclusion), transactions and isolation levels, migrations (Alembic, Prisma, Flyway, expand-contract, backfilling), or SQL vs NoSQL (MongoDB, DynamoDB, Cassandra, graph databases). Also trigger for "design the database", "model this domain", "which database should I use", or writing ORM models and migration files for PostgreSQL, MySQL, SQLite, or SQL Server.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

ddd

무료

Domain-Driven Design system for software development. Use when designing new systems with DDD principles, refactoring existing codebases toward DDD, generating code scaffolding (entities, aggregates, repositories, domain events), facilitating Event Storming sessions, creating bounded context maps, or performing code reviews with a DDD lens. Covers both strategic design (bounded contexts, subdomains, context maps, ubiquitous language) and tactical design (entities, value objects, aggregates, domain services, repositories). Supports all major architecture patterns (Hexagonal/Ports & Adapters, CQRS, Event Sourcing, Clean Architecture) with language-agnostic guidance and concrete examples in Python and TypeScript.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

genai-tk

무료

Build GenAI and agentic apps with genai-tk (tclatos/genai-tk), a YAML-driven IoC toolkit over LangChain, DeepAgents, DeerFlow, and Prefect. Use this skill whenever the user mentions genai-tk, genai_tk, `cli init`, `cli agents run`, `LangchainAgent`, `create_harness`, `get_llm`, `global_config()`, `model_id@provider`, `agents:`/`agent_defaults:` profiles, `RetrieverFactory`, `cli workflow`, or BAML extraction. Also trigger for app_conf.yaml, GENAITK_PROFILE, CodeAct, OpenSandbox, decision models, or "the toolkit".

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

langchain

무료

Build AI agents with LangChain v1 and LangGraph in Python. Use this skill whenever the user mentions LangChain, LangGraph, create_agent, init_chat_model, @tool, ToolRuntime, agent middleware, human-in-the-loop, structured output, MCP tools (langchain.mcp, MCPAdapter), RAG, multi-agent handoffs or subagents, checkpointers, Deep Agents (create_deep_agent) or Managed Deep Agents (mda, define_deep_agent). Also trigger for migrating from create_react_agent, LLMChain, AgentExecutor or langchain-mcp-adapters, stream_events, SummarizationMiddleware, and langchain-classic.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

svngoku/coding-agents-skills122026년 10월 21일 업데이트

svngoku의 스킬 모두 보기

이 스킬의 문제 신고

설치

명령으로 설치

npx skills add https://github.com/svngoku/coding-agents-skills --skill unsloth-hf-jobs

터미널에 붙여 넣고 실행합니다(Node.js 필요). Claude Code, Codex 등 설치할 에이전트를 고를 수 있습니다.

에이전트에게 부탁하기

다음 Agent Skill을 설치해 주세요.
https://github.com/svngoku/coding-agents-skills/tree/main/skills/unsloth-hf-jobs

먼저 SKILL.md와 첨부 파일의 내용을 확인해 수상한 처리(외부 전송, 인증 정보 읽기 등)가 없는지 확인한 뒤 Claude Code는 ~/.claude/skills/unsloth-hf-jobs/, Codex는 ~/.agents/skills/unsloth-hf-jobs/ 에 배치해 주세요.

이 문장을 그대로 Claude Code나 Codex에 보내면 내용을 확인한 뒤 설치해 줍니다.

GitHub에서 소스 보기

저장소: svngoku/coding-agents-skills
스타: 12
포크: 1
카테고리: 인프라·DevOps
라이선스: MIT
원문 언어: 영어
최종 업데이트: 2026년 10월 21일

스킬은 에이전트를 위한 지시와 스크립트의 묶음입니다. cccskills는 내용의 안전성을 보증하지 않습니다. 설치하기 전에 내용을 확인하세요.