[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6P855B9QA0B5DA89CEZW23B":3,"skill-same:01M6P855B9QA0B5DA89CEZW23B":71},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":38,"linkOnly":39,"markdownContent":40,"markdownBody":41,"frontMatter":42,"refName":28,"commitSha":52,"skillDirectory":53,"forksCount":13,"repositoryHtmlUrl":54,"repositoryHomepageUrl":6,"ownerHtmlUrl":55,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":56,"install":57,"resources":70},"01M6P855B9QA0B5DA89CEZW23B","llm-fine-tuning",null,"Set up infrastructure for fine-tuning LLMs with QLoRA, LoRA, and full fine-tuning using Hugging Face TRL, Axolotl, and distributed training with DeepSpeed or FSDP.","duc-duy-1612\u002FOnline-Delivery-Systems-Web","duc-duy-1612","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F192231384?v=4","01M6NSDKQQDKVTHZ8DZMZ1VYY6","01M6NSDKW8DDVTQHBT1HBDDJPT",0,"MIT","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fllm-fine-tuning\u002FSKILL.md",[],[18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37],"llm fine tuning","tuning","fine tuning","up","relevant","FSDP","scripts","critical","os","full","main","community","llms","only","hugging face","access","devops-security","SET","training","security","2026-11-10 10:12:30",false,"---\nname: llm-fine-tuning\ndescription: Set up infrastructure for fine-tuning LLMs with QLoRA, LoRA, and full\n  fine-tuning using Hugging Face TRL, Axolotl, and distributed training with DeepSpeed\n  or FSDP.\ncategory: devops\nrisk: critical\nsource: https:\u002F\u002Fgithub.com\u002FBagelHole\u002FDevOps-Security-Agent-Skills\nsource_repo: BagelHole\u002FDevOps-Security-Agent-Skills\nsource_type: community\ndate_added: '2026-09-20'\nlicense: MIT\nlicense_source: https:\u002F\u002Fgithub.com\u002FBagelHole\u002FDevOps-Security-Agent-Skills\u002Fblob\u002Fmain\u002FLICENSE\ncompatibility: Requires the relevant OS\u002Fplatform tooling and privileged access where\n  noted. Docs-only; helper scripts and templates not bundled.\nmetadata:\n  author: devops-skills\n  version: '1.0'\n---\n\n# LLM Fine-Tuning Infrastructure\n\nTrain and fine-tune open-source LLMs efficiently — from LoRA on a single GPU to distributed full fine-tuning across multi-node clusters.\n\n## When to Use This Skill\n\nUse this skill when:\n- Fine-tuning an LLM on domain-specific data (legal, medical, code, support)\n- Running QLoRA to fine-tune 70B models on consumer GPUs\n- Setting up distributed training with DeepSpeed or FSDP\n- Exporting fine-tuned adapters for production serving\n- Implementing RLHF, DPO, or instruction tuning pipelines\n\n## Prerequisites\n\n- NVIDIA GPU(s) with 24GB+ VRAM (RTX 4090 \u002F A100 \u002F H100)\n- CUDA 12.1+ and `nvidia-smi` working\n- Python 3.10+ with `pip`\n- Hugging Face account and `HF_TOKEN` for gated models\n- 500GB+ disk for model weights and training data\n\n## Quick Start: QLoRA Fine-Tuning\n\n```bash\npip install transformers datasets trl peft bitsandbytes accelerate\n\npython - \u003C\u003C'EOF'\nfrom datasets import load_dataset\nfrom transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\nfrom peft import LoraConfig, get_peft_model\nfrom trl import SFTTrainer, SFTConfig\nimport torch\n\nmodel_id = \"meta-llama\u002FLlama-3.1-8B-Instruct\"\n\n# 4-bit quantization (QLoRA)\nbnb_config = BitsAndBytesConfig(\n    load_in_4bit=True,\n    bnb_4bit_quant_type=\"nf4\",\n    bnb_4bit_compute_dtype=torch.bfloat16,\n    bnb_4bit_use_double_quant=True,\n)\n\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id, quantization_config=bnb_config, device_map=\"auto\"\n)\ntokenizer = AutoTokenizer.from_pretrained(model_id)\n\n# LoRA configuration\npeft_config = LoraConfig(\n    r=16,                    # rank\n    lora_alpha=32,\n    target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n                    \"gate_proj\", \"up_proj\", \"down_proj\"],\n    lora_dropout=0.05,\n    bias=\"none\",\n    task_type=\"CAUSAL_LM\",\n)\n\ndataset = load_dataset(\"your-org\u002Fyour-dataset\", split=\"train\")\n\ntrainer = SFTTrainer(\n    model=model,\n    args=SFTConfig(\n        output_dir=\".\u002Foutput\",\n        num_train_epochs=3,\n        per_device_train_batch_size=2,\n        gradient_accumulation_steps=8,\n        learning_rate=2e-4,\n        bf16=True,\n        logging_steps=10,\n        save_strategy=\"epoch\",\n        report_to=\"wandb\",\n    ),\n    train_dataset=dataset,\n    peft_config=peft_config,\n    processing_class=tokenizer,\n)\ntrainer.train()\ntrainer.save_model(\".\u002Ffine-tuned-model\")\nEOF\n```\n\n## Axolotl (Production Fine-Tuning Framework)\n\n```yaml\n# config.yaml — Axolotl QLoRA config for Llama 3.1\nbase_model: meta-llama\u002FLlama-3.1-8B-Instruct\nmodel_type: LlamaForCausalLM\ntokenizer_type: PreTrainedTokenizerFast\n\nload_in_4bit: true\nadapter: qlora\nlora_r: 32\nlora_alpha: 64\nlora_dropout: 0.05\nlora_target_modules:\n  - q_proj\n  - k_proj\n  - v_proj\n  - o_proj\n  - gate_proj\n  - up_proj\n  - down_proj\n\ndatasets:\n  - path: your-org\u002Fyour-dataset\n    type: alpaca              # or sharegpt, chat_template, etc.\n\ndataset_prepared_path: .\u002Fprepared-data\nval_set_size: 0.05\noutput_dir: .\u002Foutput\n\nsequence_len: 4096\nsample_packing: true         # pack multiple short samples for efficiency\n\nmicro_batch_size: 2\ngradient_accumulation_steps: 8\nnum_epochs: 3\nlearning_rate: 2e-4\noptimizer: adamw_bnb_8bit\nlr_scheduler: cosine\nwarmup_ratio: 0.05\n\nbf16: true\nflash_attention: true\n\nlogging_steps: 10\neval_steps: 100\nsave_steps: 200\nwandb_project: my-fine-tune\n```\n\n```bash\n# Run with Axolotl\npip install axolotl[flash-attn,deepspeed]\naccelerate launch -m axolotl.cli.train config.yaml\n```\n\n## Distributed Training with DeepSpeed\n\n```json\n\u002F\u002F deepspeed_zero3.json — ZeRO Stage 3 (split optimizer + gradients + params)\n{\n  \"zero_optimization\": {\n    \"stage\": 3,\n    \"offload_optimizer\": {\"device\": \"cpu\", \"pin_memory\": true},\n    \"offload_param\": {\"device\": \"cpu\", \"pin_memory\": true},\n    \"overlap_comm\": true,\n    \"contiguous_gradients\": true,\n    \"sub_group_size\": 1e9,\n    \"reduce_bucket_size\": \"auto\",\n    \"stage3_prefetch_bucket_size\": \"auto\",\n    \"stage3_param_persistence_threshold\": \"auto\",\n    \"stage3_max_live_parameters\": 1e9,\n    \"stage3_max_reuse_distance\": 1e9,\n    \"gather_16bit_weights_on_model_save\": true\n  },\n  \"bf16\": {\"enabled\": true},\n  \"gradient_clipping\": 1.0,\n  \"train_batch_size\": \"auto\",\n  \"train_micro_batch_size_per_gpu\": \"auto\"\n}\n```\n\n```bash\n# Launch 4-GPU DeepSpeed training\ndeepspeed --num_gpus=4 train.py \\\n  --deepspeed deepspeed_zero3.json \\\n  --model_name meta-llama\u002FLlama-3.1-70B-Instruct \\\n  --output_dir .\u002Foutput\n```\n\n## DPO \u002F RLHF Alignment\n\n```python\nfrom trl import DPOTrainer, DPOConfig\nfrom datasets import load_dataset\n\n# Dataset format: {\"prompt\": ..., \"chosen\": ..., \"rejected\": ...}\ndataset = load_dataset(\"your-org\u002Fpreference-data\")\n\ntrainer = DPOTrainer(\n    model=model,\n    ref_model=None,           # None = implicit reference with peft\n    args=DPOConfig(\n        output_dir=\".\u002Fdpo-output\",\n        beta=0.1,             # KL divergence weight\n        num_train_epochs=1,\n        per_device_train_batch_size=1,\n        gradient_accumulation_steps=16,\n        learning_rate=5e-7,\n        bf16=True,\n    ),\n    train_dataset=dataset[\"train\"],\n    peft_config=peft_config,\n    processing_class=tokenizer,\n)\ntrainer.train()\n```\n\n## Merging LoRA Adapters for Deployment\n\n```python\nfrom peft import PeftModel\nfrom transformers import AutoModelForCausalLM\n\n# Load base model in full precision\nbase_model = AutoModelForCausalLM.from_pretrained(\n    \"meta-llama\u002FLlama-3.1-8B-Instruct\",\n    torch_dtype=torch.bfloat16,\n    device_map=\"cpu\",\n)\n\n# Load and merge LoRA adapter\nmodel = PeftModel.from_pretrained(base_model, \".\u002Ffine-tuned-model\")\nmerged_model = model.merge_and_unload()\n\n# Save merged model (ready for vLLM serving)\nmerged_model.save_pretrained(\".\u002Fmerged-model\", safe_serialization=True)\ntokenizer.save_pretrained(\".\u002Fmerged-model\")\n\n# Push to Hugging Face Hub\nmerged_model.push_to_hub(\"your-org\u002Fyour-fine-tuned-model\")\n```\n\n## Kubernetes Training Job\n\n```yaml\napiVersion: batch\u002Fv1\nkind: Job\nmetadata:\n  name: llm-fine-tune\nspec:\n  template:\n    spec:\n      restartPolicy: OnFailure\n      nodeSelector:\n        nvidia.com\u002Fgpu.product: A100-SXM4-80GB\n      containers:\n      - name: trainer\n        image: nvcr.io\u002Fnvidia\u002Fpytorch:24.05-py3\n        command: [\"accelerate\", \"launch\", \"-m\", \"axolotl.cli.train\", \"\u002Fconfig\u002Fconfig.yaml\"]\n        resources:\n          limits:\n            nvidia.com\u002Fgpu: \"4\"\n            memory: \"320Gi\"\n          requests:\n            nvidia.com\u002Fgpu: \"4\"\n        volumeMounts:\n        - name: config\n          mountPath: \u002Fconfig\n        - name: model-cache\n          mountPath: \u002Froot\u002F.cache\u002Fhuggingface\n        - name: output\n          mountPath: \u002Foutput\n        env:\n        - name: HUGGING_FACE_HUB_TOKEN\n          valueFrom:\n            secretKeyRef:\n              name: hf-token\n              key: token\n        - name: WANDB_API_KEY\n          valueFrom:\n            secretKeyRef:\n              name: wandb-token\n              key: key\n      volumes:\n      - name: config\n        configMap:\n          name: axolotl-config\n      - name: model-cache\n        persistentVolumeClaim:\n          claimName: model-cache-pvc\n      - name: output\n        persistentVolumeClaim:\n          claimName: training-output-pvc\n```\n\n## Common Issues\n\n| Issue | Cause | Fix |\n|-------|-------|-----|\n| `CUDA out of memory` | Batch too large | Reduce `micro_batch_size`; increase `gradient_accumulation_steps` |\n| Training loss NaN | Learning rate too high | Lower LR to `1e-4` or `5e-5`; add warmup |\n| Slow training | No Flash Attention | Install `flash-attn`; enable `flash_attention: true` |\n| Poor fine-tune quality | Bad data formatting | Validate dataset format; check `sample_packing` compatibility |\n| Adapter merge errors | Mixed quantization | Merge in bf16 on CPU, not in 4-bit |\n\n## Best Practices\n\n- Use Flash Attention 2 — it's 2–4× faster and uses less memory.\n- Monitor training loss\u002Feval loss via W&B or MLflow; overfit = more dropout or less data.\n- Validate with a held-out eval set (5–10%); MMLU or custom evals for quality gates.\n- Start with LoRA r=16 before increasing — higher rank = more parameters, diminishing returns.\n- Use `sample_packing` in Axolotl to maximize GPU utilization on short sequences.\n\n## Related Skills\n\n- vllm-server (`vllm-server`) - Serve fine-tuned models\n- gpu-server-management (`gpu-server-management`) - GPU setup\n- llm-inference-scaling (`llm-inference-scaling`) - Deploy at scale\n- ai-pipeline-orchestration (`ai-pipeline-orchestration`) - Training pipelines\n\n## Limitations\n\n- Infrastructure commands can disrupt services: confirm target host\u002Fscope and have backups\u002Fsnapshots before mutating state.\n- Docs-only import: upstream scripts and templates not bundled.\n\n","# LLM Fine-Tuning Infrastructure\n\nTrain and fine-tune open-source LLMs efficiently — from LoRA on a single GPU to distributed full fine-tuning across multi-node clusters.\n\n## When to Use This Skill\n\nUse this skill when:\n- Fine-tuning an LLM on domain-specific data (legal, medical, code, support)\n- Running QLoRA to fine-tune 70B models on consumer GPUs\n- Setting up distributed training with DeepSpeed or FSDP\n- Exporting fine-tuned adapters for production serving\n- Implementing RLHF, DPO, or instruction tuning pipelines\n\n## Prerequisites\n\n- NVIDIA GPU(s) with 24GB+ VRAM (RTX 4090 \u002F A100 \u002F H100)\n- CUDA 12.1+ and `nvidia-smi` working\n- Python 3.10+ with `pip`\n- Hugging Face account and `HF_TOKEN` for gated models\n- 500GB+ disk for model weights and training data\n\n## Quick Start: QLoRA Fine-Tuning\n\n```bash\npip install transformers datasets trl peft bitsandbytes accelerate\n\npython - \u003C\u003C'EOF'\nfrom datasets import load_dataset\nfrom transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig\nfrom peft import LoraConfig, get_peft_model\nfrom trl import SFTTrainer, SFTConfig\nimport torch\n\nmodel_id = \"meta-llama\u002FLlama-3.1-8B-Instruct\"\n\n# 4-bit quantization (QLoRA)\nbnb_config = BitsAndBytesConfig(\n    load_in_4bit=True,\n    bnb_4bit_quant_type=\"nf4\",\n    bnb_4bit_compute_dtype=torch.bfloat16,\n    bnb_4bit_use_double_quant=True,\n)\n\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_id, quantization_config=bnb_config, device_map=\"auto\"\n)\ntokenizer = AutoTokenizer.from_pretrained(model_id)\n\n# LoRA configuration\npeft_config = LoraConfig(\n    r=16,                    # rank\n    lora_alpha=32,\n    target_modules=[\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\",\n                    \"gate_proj\", \"up_proj\", \"down_proj\"],\n    lora_dropout=0.05,\n    bias=\"none\",\n    task_type=\"CAUSAL_LM\",\n)\n\ndataset = load_dataset(\"your-org\u002Fyour-dataset\", split=\"train\")\n\ntrainer = SFTTrainer(\n    model=model,\n    args=SFTConfig(\n        output_dir=\".\u002Foutput\",\n        num_train_epochs=3,\n        per_device_train_batch_size=2,\n        gradient_accumulation_steps=8,\n        learning_rate=2e-4,\n        bf16=True,\n        logging_steps=10,\n        save_strategy=\"epoch\",\n        report_to=\"wandb\",\n    ),\n    train_dataset=dataset,\n    peft_config=peft_config,\n    processing_class=tokenizer,\n)\ntrainer.train()\ntrainer.save_model(\".\u002Ffine-tuned-model\")\nEOF\n```\n\n## Axolotl (Production Fine-Tuning Framework)\n\n```yaml\n# config.yaml — Axolotl QLoRA config for Llama 3.1\nbase_model: meta-llama\u002FLlama-3.1-8B-Instruct\nmodel_type: LlamaForCausalLM\ntokenizer_type: PreTrainedTokenizerFast\n\nload_in_4bit: true\nadapter: qlora\nlora_r: 32\nlora_alpha: 64\nlora_dropout: 0.05\nlora_target_modules:\n  - q_proj\n  - k_proj\n  - v_proj\n  - o_proj\n  - gate_proj\n  - up_proj\n  - down_proj\n\ndatasets:\n  - path: your-org\u002Fyour-dataset\n    type: alpaca              # or sharegpt, chat_template, etc.\n\ndataset_prepared_path: .\u002Fprepared-data\nval_set_size: 0.05\noutput_dir: .\u002Foutput\n\nsequence_len: 4096\nsample_packing: true         # pack multiple short samples for efficiency\n\nmicro_batch_size: 2\ngradient_accumulation_steps: 8\nnum_epochs: 3\nlearning_rate: 2e-4\noptimizer: adamw_bnb_8bit\nlr_scheduler: cosine\nwarmup_ratio: 0.05\n\nbf16: true\nflash_attention: true\n\nlogging_steps: 10\neval_steps: 100\nsave_steps: 200\nwandb_project: my-fine-tune\n```\n\n```bash\n# Run with Axolotl\npip install axolotl[flash-attn,deepspeed]\naccelerate launch -m axolotl.cli.train config.yaml\n```\n\n## Distributed Training with DeepSpeed\n\n```json\n\u002F\u002F deepspeed_zero3.json — ZeRO Stage 3 (split optimizer + gradients + params)\n{\n  \"zero_optimization\": {\n    \"stage\": 3,\n    \"offload_optimizer\": {\"device\": \"cpu\", \"pin_memory\": true},\n    \"offload_param\": {\"device\": \"cpu\", \"pin_memory\": true},\n    \"overlap_comm\": true,\n    \"contiguous_gradients\": true,\n    \"sub_group_size\": 1e9,\n    \"reduce_bucket_size\": \"auto\",\n    \"stage3_prefetch_bucket_size\": \"auto\",\n    \"stage3_param_persistence_threshold\": \"auto\",\n    \"stage3_max_live_parameters\": 1e9,\n    \"stage3_max_reuse_distance\": 1e9,\n    \"gather_16bit_weights_on_model_save\": true\n  },\n  \"bf16\": {\"enabled\": true},\n  \"gradient_clipping\": 1.0,\n  \"train_batch_size\": \"auto\",\n  \"train_micro_batch_size_per_gpu\": \"auto\"\n}\n```\n\n```bash\n# Launch 4-GPU DeepSpeed training\ndeepspeed --num_gpus=4 train.py \\\n  --deepspeed deepspeed_zero3.json \\\n  --model_name meta-llama\u002FLlama-3.1-70B-Instruct \\\n  --output_dir .\u002Foutput\n```\n\n## DPO \u002F RLHF Alignment\n\n```python\nfrom trl import DPOTrainer, DPOConfig\nfrom datasets import load_dataset\n\n# Dataset format: {\"prompt\": ..., \"chosen\": ..., \"rejected\": ...}\ndataset = load_dataset(\"your-org\u002Fpreference-data\")\n\ntrainer = DPOTrainer(\n    model=model,\n    ref_model=None,           # None = implicit reference with peft\n    args=DPOConfig(\n        output_dir=\".\u002Fdpo-output\",\n        beta=0.1,             # KL divergence weight\n        num_train_epochs=1,\n        per_device_train_batch_size=1,\n        gradient_accumulation_steps=16,\n        learning_rate=5e-7,\n        bf16=True,\n    ),\n    train_dataset=dataset[\"train\"],\n    peft_config=peft_config,\n    processing_class=tokenizer,\n)\ntrainer.train()\n```\n\n## Merging LoRA Adapters for Deployment\n\n```python\nfrom peft import PeftModel\nfrom transformers import AutoModelForCausalLM\n\n# Load base model in full precision\nbase_model = AutoModelForCausalLM.from_pretrained(\n    \"meta-llama\u002FLlama-3.1-8B-Instruct\",\n    torch_dtype=torch.bfloat16,\n    device_map=\"cpu\",\n)\n\n# Load and merge LoRA adapter\nmodel = PeftModel.from_pretrained(base_model, \".\u002Ffine-tuned-model\")\nmerged_model = model.merge_and_unload()\n\n# Save merged model (ready for vLLM serving)\nmerged_model.save_pretrained(\".\u002Fmerged-model\", safe_serialization=True)\ntokenizer.save_pretrained(\".\u002Fmerged-model\")\n\n# Push to Hugging Face Hub\nmerged_model.push_to_hub(\"your-org\u002Fyour-fine-tuned-model\")\n```\n\n## Kubernetes Training Job\n\n```yaml\napiVersion: batch\u002Fv1\nkind: Job\nmetadata:\n  name: llm-fine-tune\nspec:\n  template:\n    spec:\n      restartPolicy: OnFailure\n      nodeSelector:\n        nvidia.com\u002Fgpu.product: A100-SXM4-80GB\n      containers:\n      - name: trainer\n        image: nvcr.io\u002Fnvidia\u002Fpytorch:24.05-py3\n        command: [\"accelerate\", \"launch\", \"-m\", \"axolotl.cli.train\", \"\u002Fconfig\u002Fconfig.yaml\"]\n        resources:\n          limits:\n            nvidia.com\u002Fgpu: \"4\"\n            memory: \"320Gi\"\n          requests:\n            nvidia.com\u002Fgpu: \"4\"\n        volumeMounts:\n        - name: config\n          mountPath: \u002Fconfig\n        - name: model-cache\n          mountPath: \u002Froot\u002F.cache\u002Fhuggingface\n        - name: output\n          mountPath: \u002Foutput\n        env:\n        - name: HUGGING_FACE_HUB_TOKEN\n          valueFrom:\n            secretKeyRef:\n              name: hf-token\n              key: token\n        - name: WANDB_API_KEY\n          valueFrom:\n            secretKeyRef:\n              name: wandb-token\n              key: key\n      volumes:\n      - name: config\n        configMap:\n          name: axolotl-config\n      - name: model-cache\n        persistentVolumeClaim:\n          claimName: model-cache-pvc\n      - name: output\n        persistentVolumeClaim:\n          claimName: training-output-pvc\n```\n\n## Common Issues\n\n| Issue | Cause | Fix |\n|-------|-------|-----|\n| `CUDA out of memory` | Batch too large | Reduce `micro_batch_size`; increase `gradient_accumulation_steps` |\n| Training loss NaN | Learning rate too high | Lower LR to `1e-4` or `5e-5`; add warmup |\n| Slow training | No Flash Attention | Install `flash-attn`; enable `flash_attention: true` |\n| Poor fine-tune quality | Bad data formatting | Validate dataset format; check `sample_packing` compatibility |\n| Adapter merge errors | Mixed quantization | Merge in bf16 on CPU, not in 4-bit |\n\n## Best Practices\n\n- Use Flash Attention 2 — it's 2–4× faster and uses less memory.\n- Monitor training loss\u002Feval loss via W&B or MLflow; overfit = more dropout or less data.\n- Validate with a held-out eval set (5–10%); MMLU or custom evals for quality gates.\n- Start with LoRA r=16 before increasing — higher rank = more parameters, diminishing returns.\n- Use `sample_packing` in Axolotl to maximize GPU utilization on short sequences.\n\n## Related Skills\n\n- vllm-server (`vllm-server`) - Serve fine-tuned models\n- gpu-server-management (`gpu-server-management`) - GPU setup\n- llm-inference-scaling (`llm-inference-scaling`) - Deploy at scale\n- ai-pipeline-orchestration (`ai-pipeline-orchestration`) - Training pipelines\n\n## Limitations\n\n- Infrastructure commands can disrupt services: confirm target host\u002Fscope and have backups\u002Fsnapshots before mutating state.\n- Docs-only import: upstream scripts and templates not bundled.\n\n",{"name":5,"risk":25,"source":43,"license":14,"category":44,"metadata":45,"date_added":48,"description":7,"source_repo":49,"source_type":29,"compatibility":50,"license_source":51},"https:\u002F\u002Fgithub.com\u002FBagelHole\u002FDevOps-Security-Agent-Skills","devops",{"author":46,"version":47},"devops-skills","1.0","2026-09-20","BagelHole\u002FDevOps-Security-Agent-Skills","Requires the relevant OS\u002Fplatform tooling and privileged access where noted. Docs-only; helper scripts and templates not bundled.","https:\u002F\u002Fgithub.com\u002FBagelHole\u002FDevOps-Security-Agent-Skills\u002Fblob\u002Fmain\u002FLICENSE","f3312ba749561b6159f477dbfc9313c875d40a56",".agents\u002Fskills_archive\u002Fllm-fine-tuning","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612","2026-10-25 23:50:58",{"source":8,"skillName":5,"commands":58,"prompt":68,"zipAvailable":69},[59,62,65],{"tool":60,"command":61},"npx","npx skills add duc-duy-1612\u002FOnline-Delivery-Systems-Web --skill llm-fine-tuning",{"tool":63,"command":64},"bunx","bunx skills add duc-duy-1612\u002FOnline-Delivery-Systems-Web --skill llm-fine-tuning",{"tool":66,"command":67},"pnpm","pnpm dlx skills add duc-duy-1612\u002FOnline-Delivery-Systems-Web --skill llm-fine-tuning","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fllm-fine-tuning\u002FSKILL.md",true,[],[72,95,119,144,172,196],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":83,"publishedAt":38},"01M6Q0FA5FZ410Z33Z3TFQ4DBA","zero-trust","Implement zero-trust network architecture. Configure identity-based access, micro-segmentation, and continuous verification. Use when implementing modern security architectures.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fzero-trust\u002FSKILL.md",[78],{"categoryNo":79,"categoryName":80,"taxonomyType":81,"taxonomyTypeDesc":82},"01JBSEEDCAT000000000000011","运维与部署","SCENARIO","使用场景",[84,85,86,87,88,89,74,90,34,33,91,31,29,28,92,25,93,24,22,94],"ZeRO","trust","zero trust","architecture","assessment","verification","segmentation","configure","active","scope","implementing",{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":38},"01M6Q0F1YZJW3KX5FBKEVZ8CV4","zcode-delegate","Delegate coding tasks to the Z.AI ZCode CLI only when the user explicitly requests it, while the orchestrator retains review and landing responsibility.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fzcode-delegate\u002FSKILL.md",[],[102,103,104,105,106,107,108,109,110,111,112,113,114,31,115,29,116,117,25,118],"delegate","zcode","zcode delegate","review","cli","orchestration","tasks","run","Bash","model","node","read","orchestrator","files","desktop app","app","agent-orchestration",{"skillNo":120,"name":121,"nameZh":6,"description":122,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":38},"01M6Q0ETK41VGNMYEXW0T00AJN","youtube-transcript","Fetch YouTube transcripts through DeepAPI or local fallback tooling and save clean text output.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fyoutube-transcript\u002FSKILL.md",[],[126,127,128,129,130,131,132,133,134,135,136,137,138,29,28,139,140,141,142,143],"transcript","youtube transcript","youtube","safe","github","codex","research","license","https","text","fetch","tooling","clean","local","output","deepapi","youtube transcripts","blob",{"skillNo":145,"name":146,"nameZh":6,"description":147,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":148,"categories":149,"keywords":161,"publishedAt":38},"01M6Q0EPPH2PVMDW40017F8WX0","youtube-transcript-skills","Fetch YouTube video transcripts, search videos\u002Fchannels, browse channels, and extract playlists via the getyoutubetranscript.com API - free tier, no card required.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fyoutube-transcript-skills\u002FSKILL.md",[150,153,156],{"categoryNo":151,"categoryName":152,"taxonomyType":81,"taxonomyTypeDesc":82},"01JBSEEDCAT000000000000010","后端与接口",{"categoryNo":154,"categoryName":155,"taxonomyType":81,"taxonomyTypeDesc":82},"01JBSEEDCAT000000000000016","图像与多媒体",{"categoryNo":157,"categoryName":158,"taxonomyType":159,"taxonomyTypeDesc":160},"01JBSEEDCAT000000000000030","开发工程师","PROFESSION","适用职业",[126,127,128,162,129,163,130,131,133,164,165,166,167,136,168,29,28,169,170,171],"youtube transcript skills","gemini","search","api","Video","browse","api-integration","free","card","channels",{"skillNo":173,"name":174,"nameZh":6,"description":175,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":176,"categories":177,"keywords":182,"publishedAt":38},"01M6Q0EAB96PX83W5PTE8A0R8D","youtube-notetaker","Turn YouTube talks into local study notes with slides, transcripts, editable annotations, and a markdown-backed viewer.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fyoutube-notetaker\u002FSKILL.md",[178,181],{"categoryNo":179,"categoryName":180,"taxonomyType":81,"taxonomyTypeDesc":82},"01JBSEEDCAT000000000000003","演示文稿",{"categoryNo":154,"categoryName":155,"taxonomyType":81,"taxonomyTypeDesc":82},[128,183,184,106,129,185,186,187,188,133,166,189,190,191,192,28,193,139,194,195],"youtube notetaker","notetaker","workflow","official","notes","code","markdown","slides","annotations","README","study","dair-academy","viewer",{"skillNo":197,"name":198,"nameZh":6,"description":199,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":200,"categories":201,"keywords":203,"publishedAt":38},"01M6Q0E55F0GW5MX48V8HXS8B5","youtube-full","Fetch YouTube transcripts, search videos, browse channels, and extract playlists via TranscriptAPI — no yt-dlp, no Google API key, works from any cloud server.","https:\u002F\u002Fgithub.com\u002Fduc-duy-1612\u002FOnline-Delivery-Systems-Web\u002Fblob\u002Fmain\u002F.agents\u002Fskills_archive\u002Fyoutube-full\u002FSKILL.md",[202],{"categoryNo":151,"categoryName":152,"taxonomyType":81,"taxonomyTypeDesc":82},[204,128,27,205,129,163,164,206,207,165,166,167,136,208,168,29,28,169,209,142],"youtube full","google","server","dlp","credits","automatic"]