[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6DX8Y8PX49YR6C6SHVB3JY0":3,"skill-same:01M6DX8Y8PX49YR6C6SHVB3JY0":87},{"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":39,"linkOnly":40,"markdownContent":41,"markdownBody":42,"frontMatter":43,"refName":64,"commitSha":65,"skillDirectory":66,"forksCount":67,"repositoryHtmlUrl":68,"repositoryHomepageUrl":6,"ownerHtmlUrl":69,"ownerBlogUrl":70,"ownerDescription":71,"updatedAt":72,"install":73,"resources":86},"01M6DX8Y8PX49YR6C6SHVB3JY0","trl-training",null,"Train and fine-tune transformer language models using TRL (Transformers Reinforcement Learning). Supports SFT, DPO, GRPO, KTO, RLOO and Reward Model training via CLI commands.","cianfhoghlaim\u002Fcianfhoghlaim","cianfhoghlaim","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F57019547?v=4","01M6DWR3GAV495PYDTZSR1FZNX","Cian Mac (an Déisigh Uí) Liatháin || Deacy-Lyons","01M6DWR3N3Z768BB64XVJN0AY2",4,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fhuggingface\u002Ftrl-training\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38],"training","trl training","tuning","cli","via","reinforcement","commands","models","en","model","learning","Fine-Tuning","machine-learning","Transformers","RLOO","RLHF","HuggingFace","GRPO","KTO","DPO","2026-11-10 10:12:30",false,"---\nname: trl-training\ndescription: Train and fine-tune transformer language models using TRL (Transformers Reinforcement Learning). Supports SFT, DPO, GRPO, KTO, RLOO and Reward Model training via CLI commands.\nlicense: Apache-2.0\nmetadata:\n  version: \"1.0.0\"\n  author: huggingface\n  commands:\n    - trl sft\n    - trl dpo\n    - trl grpo\n    - trl kto\n    - trl rloo\n    - trl reward\n  categories:\n    - machine-learning\n    - llm-training\n    - reinforcement-learning\n  tags:\n    - rlhf\n    - supervised-fine-tuning\n    - dpo\n    - grpo\n    - huggingface\n    - transformers\n  documentation: https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftrl\u002Fen\u002Fclis\n---\n\n# TRL Training Skill\n\nYou are an expert at using the TRL (Transformers Reinforcement Learning) library to train and fine-tune large language models.\n\n## Overview\n\nTRL provides CLI commands for post-training foundation models using state-of-the-art techniques:\n\n- **SFT** (Supervised Fine-Tuning): Fine-tune models on instruction-following or conversational datasets\n- **DPO** (Direct Preference Optimization): Align models using preference data\n- **GRPO** (Group Relative Policy Optimization): Train models by ranking multiple sampled outputs relative to each other and optimizing based on their comparative rewards.\n- **RLOO** (Reinforce Leave One Out): Online RL training with generation-based rewards\n- **Reward Model Training**: Train reward models for RLHF\n\nTRL is built on top of Hugging Face Transformers and Accelerate, providing seamless integration with the Hugging Face ecosystem.\n\n## Core Commands\n\n### trl sft - Supervised Fine-Tuning\n\nFine-tune language models on instruction-following or conversational datasets.\n\n**Full training:**\n\n```bash\ntrl sft \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B \\\n  --dataset_name trl-lib\u002FCapybara \\\n  --learning_rate 2.0e-5 \\\n  --num_train_epochs 1 \\\n  --packing \\\n  --per_device_train_batch_size 2 \\\n  --gradient_accumulation_steps 8 \\\n  --eos_token '\u003C|im_end|>' \\\n  --eval_strategy steps \\\n  --eval_steps 100 \\\n  --output_dir Qwen2-0.5B-SFT \\\n  --push_to_hub\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl sft \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B \\\n  --dataset_name trl-lib\u002FCapybara \\\n  --learning_rate 2.0e-4 \\\n  --num_train_epochs 1 \\\n  --packing \\\n  --per_device_train_batch_size 2 \\\n  --gradient_accumulation_steps 8 \\\n  --eos_token '\u003C|im_end|>' \\\n  --eval_strategy steps \\\n  --eval_steps 100 \\\n  --use_peft \\\n  --lora_r 32 \\\n  --lora_alpha 16 \\\n  --output_dir Qwen2-0.5B-SFT \\\n  --push_to_hub\n```\n\n### trl dpo - Direct Preference Optimization\n\nAlign models using preference data (chosen\u002Frejected pairs).\n\n**Full training:**\n\n```bash\ntrl dpo \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --learning_rate 5.0e-7 \\\n  --num_train_epochs 1 \\\n  --per_device_train_batch_size 2 \\\n  --max_steps 1000 \\\n  --gradient_accumulation_steps 8 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --output_dir Qwen2-0.5B-DPO \\\n  --no_remove_unused_columns\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl dpo \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --learning_rate 5.0e-6 \\\n  --num_train_epochs 1 \\\n  --per_device_train_batch_size 2 \\\n  --max_steps 1000 \\\n  --gradient_accumulation_steps 8 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --output_dir Qwen2-0.5B-DPO \\\n  --no_remove_unused_columns \\\n  --use_peft \\\n  --lora_r 32 \\\n  --lora_alpha 16\n```\n\n### trl grpo - Group Relative Policy Optimization\n\nTrain models using reward functions or LLM-as-a-judge for evaluating generations and providing rewards.\n\n**Basic usage:**\n\n```bash\ntrl grpo \\\n  --model_name_or_path Qwen\u002FQwen2.5-0.5B \\\n  --dataset_name trl-lib\u002Fgsm8k \\\n  --reward_funcs accuracy_reward \\\n  --output_dir Qwen2-0.5B-GRPO \\\n  --push_to_hub\n```\n\n### trl rloo - Reinforce Leave One Out\n\nOnline RL training where the model generates text and receives rewards based on custom criteria.\n\n**Basic usage:**\n\n```bash\ntrl rloo \\\n  --model_name_or_path Qwen\u002FQwen2.5-0.5B \\\n  --dataset_name trl-lib\u002Ftldr \\\n  --reward_model_name_or_path sentiment-analysis:nlptown\u002Fbert-base-multilingual-uncased-sentiment \\\n  --output_dir Qwen2-0.5B-RLOO \\\n  --push_to_hub\n```\n\n### trl reward - Reward Model Training\n\nTrain a reward model to score text quality for RLHF.\n\n**Full training:**\n\n```bash\ntrl reward \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --output_dir Qwen2-0.5B-Reward \\\n  --per_device_train_batch_size 8 \\\n  --num_train_epochs 1 \\\n  --learning_rate 1.0e-5 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --max_length 2048\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl reward \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --output_dir Qwen2-0.5B-Reward-LoRA \\\n  --per_device_train_batch_size 8 \\\n  --num_train_epochs 1 \\\n  --learning_rate 1.0e-4 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --max_length 2048 \\\n  --use_peft \\\n  --lora_task_type SEQ_CLS \\\n  --lora_r 32 \\\n  --lora_alpha 16\n```\n\n## Configuration Files\n\nTRL supports YAML configuration files for reproducible training. All CLI arguments can be specified in a config file.\n\n**Example config (sft_config.yaml):**\n\n```yaml\nmodel_name_or_path: Qwen\u002FQwen2.5-0.5B\ndataset_name: trl-lib\u002FCapybara\nlearning_rate: 2.0e-5\nnum_train_epochs: 1\nper_device_train_batch_size: 8\ngradient_accumulation_steps: 2\noutput_dir: .\u002Fsft_output\nuse_peft: true\nlora_r: 16\nlora_alpha: 16\nreport_to: trackio\n```\n\n**Launch with config:**\n\n```bash\ntrl sft --config sft_config.yaml\n```\n\n**Override config values:**\n\n```bash\ntrl sft --config sft_config.yaml --learning_rate 1.0e-5\n```\n\n## Distributed Training\n\nTRL integrates with Accelerate for multi-GPU and multi-node training.\n\n**Multi-GPU training:**\n\n```bash\ntrl sft \\\n  --config sft_config.yaml \\\n  --num_processes 4\n```\n\n**Use predefined Accelerate configs:**\n\nTRL provides predefined configs: `single_gpu`, `multi_gpu`, `fsdp1`, `fsdp2`, `zero1`, `zero2`, `zero3`\n\n```bash\ntrl sft \\\n  --config sft_config.yaml \\\n  --accelerate_config zero2\n```\n\n**Custom Accelerate config:**\n\n```bash\n# Generate custom config\naccelerate config\n\n# Use custom config\ntrl sft --config sft_config.yaml --config_file ~\u002F.cache\u002Fhuggingface\u002Faccelerate\u002Fdefault_config.yaml\n```\n\n**Fully Sharded Data Parallel (FSDP):**\n\n```bash\ntrl sft --config sft_config.yaml --accelerate_config fsdp2\n```\n\n**DeepSpeed ZeRO:**\n\n```bash\ntrl sft --config sft_config.yaml --accelerate_config zero3\n```\n\n## Troubleshooting\n\n### CUDA Out of Memory\n\n- Reduce `--per_device_train_batch_size` and increase `--gradient_accumulation_steps`\n- Enable `--use_peft` for LoRA training\n- Use `--gradient_checkpointing` to save memory\n- Try smaller model or longer sequence truncation\n\n### Dataset Loading Issues\n\n- Verify dataset exists: check Hugging Face Hub or local path\n- Check dataset format matches expected columns\n- Use `--dataset_config` for multi-config datasets\n- Inspect dataset: `from datasets import load_dataset; ds = load_dataset(name)`\n\n### Model Loading Issues\n\n- Verify model exists on Hugging Face Hub\n- Check if gated model requires authentication: `hf auth login`\n- For local models, provide absolute path\n- Ensure sufficient disk space and memory\n\n### Slow Training\n\n- Enable dataset `--packing` for short sequences\n- Use larger `--per_device_train_batch_size` if memory allows\n- Enable `--tf32` for faster computation on Ampere GPUs\n- Use `--bf16` on supported hardware\n- Consider multi-GPU training with `--num_processes`\n\n### Generation Issues (GRPO\u002FRLOO)\n\n- Check prompt format in dataset\n- Adjust `--temperature` and `--top_p` for generation\n- Verify the reward function (for GRPO\u002FRLOO)\n\n## Additional Resources\n\n- **Documentation**: https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftrl\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftrl\n- **Examples**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftrl\u002Ftree\u002Fmain\u002Fexamples\n\n## Best Practices\n\n1. **Start with SFT**: Always fine-tune base models with SFT before preference alignment\n2. **Use LoRA for efficiency**: Enable `--use_peft` for faster training and lower memory\n3. **Monitor training**: Use `--report_to trackio` (or `--report_to wandb` or `--report_to tensorboard`) for tracking\n4. **Save checkpoints**: TRL automatically saves checkpoints in `--output_dir`\n5. **Test on small datasets first**: Verify pipeline works before full training\n6. **Use configuration files**: Create YAML configs for reproducibility\n7. **Leverage Accelerate**: Use multi-GPU training for faster iteration\n\nWhen helping users with TRL:\n- Always check which training method is appropriate for their use case\n- Verify dataset format matches the expected schema\n- Recommend starting with smaller models for testing\n- Suggest LoRA for resource-constrained environments\n- Point to specific documentation sections for advanced features\n","# TRL Training Skill\n\nYou are an expert at using the TRL (Transformers Reinforcement Learning) library to train and fine-tune large language models.\n\n## Overview\n\nTRL provides CLI commands for post-training foundation models using state-of-the-art techniques:\n\n- **SFT** (Supervised Fine-Tuning): Fine-tune models on instruction-following or conversational datasets\n- **DPO** (Direct Preference Optimization): Align models using preference data\n- **GRPO** (Group Relative Policy Optimization): Train models by ranking multiple sampled outputs relative to each other and optimizing based on their comparative rewards.\n- **RLOO** (Reinforce Leave One Out): Online RL training with generation-based rewards\n- **Reward Model Training**: Train reward models for RLHF\n\nTRL is built on top of Hugging Face Transformers and Accelerate, providing seamless integration with the Hugging Face ecosystem.\n\n## Core Commands\n\n### trl sft - Supervised Fine-Tuning\n\nFine-tune language models on instruction-following or conversational datasets.\n\n**Full training:**\n\n```bash\ntrl sft \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B \\\n  --dataset_name trl-lib\u002FCapybara \\\n  --learning_rate 2.0e-5 \\\n  --num_train_epochs 1 \\\n  --packing \\\n  --per_device_train_batch_size 2 \\\n  --gradient_accumulation_steps 8 \\\n  --eos_token '\u003C|im_end|>' \\\n  --eval_strategy steps \\\n  --eval_steps 100 \\\n  --output_dir Qwen2-0.5B-SFT \\\n  --push_to_hub\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl sft \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B \\\n  --dataset_name trl-lib\u002FCapybara \\\n  --learning_rate 2.0e-4 \\\n  --num_train_epochs 1 \\\n  --packing \\\n  --per_device_train_batch_size 2 \\\n  --gradient_accumulation_steps 8 \\\n  --eos_token '\u003C|im_end|>' \\\n  --eval_strategy steps \\\n  --eval_steps 100 \\\n  --use_peft \\\n  --lora_r 32 \\\n  --lora_alpha 16 \\\n  --output_dir Qwen2-0.5B-SFT \\\n  --push_to_hub\n```\n\n### trl dpo - Direct Preference Optimization\n\nAlign models using preference data (chosen\u002Frejected pairs).\n\n**Full training:**\n\n```bash\ntrl dpo \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --learning_rate 5.0e-7 \\\n  --num_train_epochs 1 \\\n  --per_device_train_batch_size 2 \\\n  --max_steps 1000 \\\n  --gradient_accumulation_steps 8 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --output_dir Qwen2-0.5B-DPO \\\n  --no_remove_unused_columns\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl dpo \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --learning_rate 5.0e-6 \\\n  --num_train_epochs 1 \\\n  --per_device_train_batch_size 2 \\\n  --max_steps 1000 \\\n  --gradient_accumulation_steps 8 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --output_dir Qwen2-0.5B-DPO \\\n  --no_remove_unused_columns \\\n  --use_peft \\\n  --lora_r 32 \\\n  --lora_alpha 16\n```\n\n### trl grpo - Group Relative Policy Optimization\n\nTrain models using reward functions or LLM-as-a-judge for evaluating generations and providing rewards.\n\n**Basic usage:**\n\n```bash\ntrl grpo \\\n  --model_name_or_path Qwen\u002FQwen2.5-0.5B \\\n  --dataset_name trl-lib\u002Fgsm8k \\\n  --reward_funcs accuracy_reward \\\n  --output_dir Qwen2-0.5B-GRPO \\\n  --push_to_hub\n```\n\n### trl rloo - Reinforce Leave One Out\n\nOnline RL training where the model generates text and receives rewards based on custom criteria.\n\n**Basic usage:**\n\n```bash\ntrl rloo \\\n  --model_name_or_path Qwen\u002FQwen2.5-0.5B \\\n  --dataset_name trl-lib\u002Ftldr \\\n  --reward_model_name_or_path sentiment-analysis:nlptown\u002Fbert-base-multilingual-uncased-sentiment \\\n  --output_dir Qwen2-0.5B-RLOO \\\n  --push_to_hub\n```\n\n### trl reward - Reward Model Training\n\nTrain a reward model to score text quality for RLHF.\n\n**Full training:**\n\n```bash\ntrl reward \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --output_dir Qwen2-0.5B-Reward \\\n  --per_device_train_batch_size 8 \\\n  --num_train_epochs 1 \\\n  --learning_rate 1.0e-5 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --max_length 2048\n```\n\n**Train with LoRA adapters:**\n\n```bash\ntrl reward \\\n  --model_name_or_path Qwen\u002FQwen2-0.5B-Instruct \\\n  --dataset_name trl-lib\u002Fultrafeedback_binarized \\\n  --output_dir Qwen2-0.5B-Reward-LoRA \\\n  --per_device_train_batch_size 8 \\\n  --num_train_epochs 1 \\\n  --learning_rate 1.0e-4 \\\n  --eval_strategy steps \\\n  --eval_steps 50 \\\n  --max_length 2048 \\\n  --use_peft \\\n  --lora_task_type SEQ_CLS \\\n  --lora_r 32 \\\n  --lora_alpha 16\n```\n\n## Configuration Files\n\nTRL supports YAML configuration files for reproducible training. All CLI arguments can be specified in a config file.\n\n**Example config (sft_config.yaml):**\n\n```yaml\nmodel_name_or_path: Qwen\u002FQwen2.5-0.5B\ndataset_name: trl-lib\u002FCapybara\nlearning_rate: 2.0e-5\nnum_train_epochs: 1\nper_device_train_batch_size: 8\ngradient_accumulation_steps: 2\noutput_dir: .\u002Fsft_output\nuse_peft: true\nlora_r: 16\nlora_alpha: 16\nreport_to: trackio\n```\n\n**Launch with config:**\n\n```bash\ntrl sft --config sft_config.yaml\n```\n\n**Override config values:**\n\n```bash\ntrl sft --config sft_config.yaml --learning_rate 1.0e-5\n```\n\n## Distributed Training\n\nTRL integrates with Accelerate for multi-GPU and multi-node training.\n\n**Multi-GPU training:**\n\n```bash\ntrl sft \\\n  --config sft_config.yaml \\\n  --num_processes 4\n```\n\n**Use predefined Accelerate configs:**\n\nTRL provides predefined configs: `single_gpu`, `multi_gpu`, `fsdp1`, `fsdp2`, `zero1`, `zero2`, `zero3`\n\n```bash\ntrl sft \\\n  --config sft_config.yaml \\\n  --accelerate_config zero2\n```\n\n**Custom Accelerate config:**\n\n```bash\n# Generate custom config\naccelerate config\n\n# Use custom config\ntrl sft --config sft_config.yaml --config_file ~\u002F.cache\u002Fhuggingface\u002Faccelerate\u002Fdefault_config.yaml\n```\n\n**Fully Sharded Data Parallel (FSDP):**\n\n```bash\ntrl sft --config sft_config.yaml --accelerate_config fsdp2\n```\n\n**DeepSpeed ZeRO:**\n\n```bash\ntrl sft --config sft_config.yaml --accelerate_config zero3\n```\n\n## Troubleshooting\n\n### CUDA Out of Memory\n\n- Reduce `--per_device_train_batch_size` and increase `--gradient_accumulation_steps`\n- Enable `--use_peft` for LoRA training\n- Use `--gradient_checkpointing` to save memory\n- Try smaller model or longer sequence truncation\n\n### Dataset Loading Issues\n\n- Verify dataset exists: check Hugging Face Hub or local path\n- Check dataset format matches expected columns\n- Use `--dataset_config` for multi-config datasets\n- Inspect dataset: `from datasets import load_dataset; ds = load_dataset(name)`\n\n### Model Loading Issues\n\n- Verify model exists on Hugging Face Hub\n- Check if gated model requires authentication: `hf auth login`\n- For local models, provide absolute path\n- Ensure sufficient disk space and memory\n\n### Slow Training\n\n- Enable dataset `--packing` for short sequences\n- Use larger `--per_device_train_batch_size` if memory allows\n- Enable `--tf32` for faster computation on Ampere GPUs\n- Use `--bf16` on supported hardware\n- Consider multi-GPU training with `--num_processes`\n\n### Generation Issues (GRPO\u002FRLOO)\n\n- Check prompt format in dataset\n- Adjust `--temperature` and `--top_p` for generation\n- Verify the reward function (for GRPO\u002FRLOO)\n\n## Additional Resources\n\n- **Documentation**: https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftrl\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftrl\n- **Examples**: https:\u002F\u002Fgithub.com\u002Fhuggingface\u002Ftrl\u002Ftree\u002Fmain\u002Fexamples\n\n## Best Practices\n\n1. **Start with SFT**: Always fine-tune base models with SFT before preference alignment\n2. **Use LoRA for efficiency**: Enable `--use_peft` for faster training and lower memory\n3. **Monitor training**: Use `--report_to trackio` (or `--report_to wandb` or `--report_to tensorboard`) for tracking\n4. **Save checkpoints**: TRL automatically saves checkpoints in `--output_dir`\n5. **Test on small datasets first**: Verify pipeline works before full training\n6. **Use configuration files**: Create YAML configs for reproducibility\n7. **Leverage Accelerate**: Use multi-GPU training for faster iteration\n\nWhen helping users with TRL:\n- Always check which training method is appropriate for their use case\n- Verify dataset format matches the expected schema\n- Recommend starting with smaller models for testing\n- Suggest LoRA for resource-constrained environments\n- Point to specific documentation sections for advanced features\n",{"name":5,"license":15,"metadata":44,"description":7},{"tags":45,"author":50,"version":52,"commands":53,"categories":60,"documentation":63},[46,47,48,49,50,51],"rlhf","supervised-fine-tuning","dpo","grpo","huggingface","transformers","1.0.0",[54,55,56,57,58,59],"trl sft","trl dpo","trl grpo","trl kto","trl rloo","trl reward",[31,61,62],"llm-training","reinforcement-learning","https:\u002F\u002Fhuggingface.co\u002Fdocs\u002Ftrl\u002Fen\u002Fclis","main","5ace051462d0639f59bcc60912f4ec08c1a96772",".agents\u002Fskills\u002Fhuggingface\u002Ftrl-training",1,"https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim","cianfhoghlaim.ie","Irish-British teacher, mathematician, foghlaimeoir || minority-language AI researcher || Lucht Siúil ||\r\nFine Gael + Alliance Party + Liberal Democrats Member","2026-10-12 08:07:09",{"source":8,"skillName":5,"commands":74,"prompt":84,"zipAvailable":85},[75,78,81],{"tool":76,"command":77},"npx","npx skills add cianfhoghlaim\u002Fcianfhoghlaim --skill trl-training",{"tool":79,"command":80},"bunx","bunx skills add cianfhoghlaim\u002Fcianfhoghlaim --skill trl-training",{"tool":82,"command":83},"pnpm","pnpm dlx skills add cianfhoghlaim\u002Fcianfhoghlaim --skill trl-training","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fhuggingface\u002Ftrl-training\u002FSKILL.md",true,[],[88,109,128,149,173,201],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":39},"01M6DXR6F7GB4JSD67XTX6WY4A","dagster-init","Initialize a dagster project using the create-dagster cli. Create a dagster project, uv virtual environment, and everything needed for a user to run dg dev or dg check defs successfully. (project)","MIT","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fdagster\u002Fdagster-dspy\u002F.claude\u002Fskills\u002Fdagster-init\u002FSKILL.md",[],[96,97,98,99,100,101,102,103,104,105,106,107,108,22],"init","dagster","dagster init","run","project","uv","environment","user","initialize","check","virtual","dev","create",{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":120,"publishedAt":39},"01M6DXQZ2HSP548JH2V3FQ7H2V","create-custom-dagster-component","Create a custom Dagster Component with demo mode support, realistic asset structure, and optional custom scaffolder using the dg CLI. Use this skill if there is no Component included in an existing integration or if Dagster does not have the integration.","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002Fdocs\u002Fresearch\u002Fdagster\u002Fdagster-dspy\u002F.claude\u002Fskills\u002Fcreate-custom-dagster-component\u002FSKILL.md",[115],{"categoryNo":116,"categoryName":117,"taxonomyType":118,"taxonomyTypeDesc":119},"01JBSEEDCAT000000000000023","Skill与智能体开发","SCENARIO","使用场景",[121,108,97,122,123,22,124,125,126,127],"component","create custom dagster component","support","demo","structure","use","integration",{"skillNo":129,"name":130,"nameZh":6,"description":131,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":132,"categories":133,"keywords":134,"publishedAt":39},"01M6DXHPZMQ4VAXH9PGBK2PNNP","pydantic-ai-harness","Extend Pydantic AI agents with batteries-included capabilities from pydantic-ai-harness -- currently Code Mode, which collapses many tool calls into one sandboxed Python execution. Use when the user mentions pydantic-ai-harness, CodeMode, Monty, code mode, or tool sandboxing, when they want an agent to run agent-written Python, or when a Pydantic AI agent would benefit from orchestrating multiple tool calls in a single sandboxed script.","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fpydantic\u002Fpydantic-ai-harness\u002FSKILL.md",[],[135,136,137,138,99,139,140,126,103,141,142,143,144,145,146,147,148],"Pydantic","harness","pydantic ai","pydantic ai harness","batteries","pydantic-ai","tool","sandboxing","execution","capabilities","code","ai agents","python","ai agent",{"skillNo":150,"name":151,"nameZh":6,"description":152,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":153,"categories":154,"keywords":159,"publishedAt":39},"01M6DXH7JE54YEHM58JY8KY2AK","building-pydantic-ai-agents","Build AI agents with Pydantic AI — tools, capabilities (including on-demand loading), structured output, streaming, testing, and multi-agent patterns. Use when the user mentions Pydantic AI, imports pydantic_ai, or asks to build an AI agent, add tools\u002Fcapabilities, defer capability loading, stream output, define agents from YAML, or test agent behavior.","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fpydantic\u002Fbuilding-pydantic-ai-agents\u002FSKILL.md",[155,158],{"categoryNo":156,"categoryName":157,"taxonomyType":118,"taxonomyTypeDesc":119},"01JBSEEDCAT000000000000008","测试与质量",{"categoryNo":116,"categoryName":117,"taxonomyType":118,"taxonomyTypeDesc":119},[146,135,137,160,147,161,148,162,163,164,165,166,126,167,168,169,170,171,144,172],"building pydantic ai agents","testing","structured output","define","agent patterns","patterns","build","behavior","multi-agent","Tools","yaml","streaming","test",{"skillNo":174,"name":175,"nameZh":6,"description":176,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":177,"categories":178,"keywords":184,"publishedAt":39},"01M6DXGCVB0WSHBQ7BSA56A12C","motherduck-share-data","Create and manage MotherDuck data shares for zero-copy, read-only data distribution. Use whenever someone wants to share a database with team members, another organization, or the public — covers CREATE SHARE, access\u002Fvisibility\u002Fupdate modes, GRANT READ ON SHARE, attaching share URLs, UPDATE SHARE, and REFRESH DATABASE.","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fmotherduck\u002Fmotherduck-share-data\u002FSKILL.md",[179],{"categoryNo":180,"categoryName":181,"taxonomyType":182,"taxonomyTypeDesc":183},"01JBSEEDCAT000000000000033","数据分析师","PROFESSION","适用职业",[185,186,187,188,126,189,190,191,192,193,194,195,196,197,198,199,200,108],"share","data","motherduck","motherduck share data","distribution","organization","manage","only","refresh","read","public","access","team","ZeRO","database","update",{"skillNo":202,"name":203,"nameZh":6,"description":204,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":92,"licenseFileUrl":6,"sourceHtmlUrl":205,"categories":206,"keywords":207,"publishedAt":39},"01M6DXG9NMFYC72KJWQVJ1D6BZ","motherduck-security-governance","Explain MotherDuck security, governance, and access-control patterns. Use for any question about SOC 2, GDPR, compliance, data residency, regions, SSO, service accounts, token handling, tenant isolation, sharing boundaries, snapshots and recovery, or governance posture — including when a security_compliance_owner, technical_owner, or application_builder is evaluating MotherDuck.","https:\u002F\u002Fgithub.com\u002Fcianfhoghlaim\u002Fcianfhoghlaim\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fmotherduck\u002Fmotherduck-security-governance\u002FSKILL.md",[],[208,209,210,187,211,212,213,196,214,215,216,217,218,186,219,220,221,222,126,223],"governance","security","motherduck security governance","sso","sharing","recovery","explain","technical","compliance","soc","gdpr","snapshots","token","control","question","builder"]