Agent SkillsRightNow-AI/openfang › llm-finetuning

llm-finetuning

GitHub

LLM微调专家技能,涵盖LoRA/QLoRA参数高效微调、数据集准备、训练优化及评估策略。指导使用Hugging Face PEFT进行模型适配、超参数调优和适配器部署。

crates/openfang-skills/bundled/llm-finetuning/SKILL.md RightNow-AI/openfang

Trigger Scenarios

需要微调大语言模型 配置LoRA或QLoRA训练 准备指令微调数据集 优化模型训练性能

Install

npx skills add RightNow-AI/openfang --skill llm-finetuning -g -y
More Options

Non-standard path

npx skills add https://github.com/RightNow-AI/openfang/tree/main/crates/openfang-skills/bundled/llm-finetuning -g -y

Use without installing

npx skills use RightNow-AI/openfang@llm-finetuning

指定 Agent (Claude Code)

npx skills add RightNow-AI/openfang --skill llm-finetuning -a claude-code -g -y

安装 repo 全部 skill

npx skills add RightNow-AI/openfang --all -g -y

预览 repo 内 skill

npx skills add RightNow-AI/openfang --list

SKILL.md

Frontmatter
{
    "name": "llm-finetuning",
    "description": "LLM fine-tuning expert for LoRA, QLoRA, dataset preparation, and training optimization"
}

LLM Fine-Tuning Expert

A deep learning specialist with hands-on expertise in fine-tuning large language models using parameter-efficient methods, dataset curation, and training optimization. This skill provides guidance for adapting foundation models to specific domains and tasks using LoRA, QLoRA, and the Hugging Face PEFT ecosystem, covering dataset preparation, hyperparameter selection, evaluation strategies, and adapter deployment.

Key Principles

  • Fine-tuning is about teaching a model your task format and domain knowledge, not about teaching it language; start with the strongest base model you can afford to run
  • Dataset quality matters far more than quantity; 1,000 carefully curated, diverse, high-quality examples often outperform 100,000 noisy ones
  • Use parameter-efficient fine-tuning (LoRA/QLoRA) to reduce memory requirements by orders of magnitude while achieving performance comparable to full fine-tuning
  • Evaluate with task-specific metrics and human review, not just perplexity; a model with lower perplexity may still produce worse outputs for your specific use case
  • Track every experiment with exact hyperparameters, dataset versions, and base model checkpoints so that results are reproducible and comparable

Techniques

  • Configure LoRA with appropriate rank (r=8 to 64), alpha (typically 2x rank), and target modules (q_proj, v_proj for attention, or all linear layers for broader adaptation)
  • Use QLoRA for memory-constrained setups: load the base model in 4-bit NormalFloat quantization, attach LoRA adapters in fp16/bf16, and train with paged optimizers to handle memory spikes
  • Format datasets as instruction-response pairs with consistent templates; include a system field for persona or context, an instruction field for the task, and a response field for the expected output
  • Apply the PEFT library workflow: load base model, create LoRA config, get_peft_model(), train with the Hugging Face Trainer or a custom loop, then save and load adapters independently
  • Set training hyperparameters carefully: learning rate between 1e-5 and 2e-4 with cosine schedule, 1-5 epochs (watch for overfitting), warmup ratio of 0.03-0.1, and gradient accumulation to simulate larger batch sizes
  • Evaluate with multiple signals: validation loss for overfitting detection, task-specific metrics (ROUGE for summarization, exact match for QA), and structured human evaluation on a held-out set

Common Patterns

  • Domain Adaptation: Fine-tune on domain-specific text (legal, medical, financial) to teach the model terminology, reasoning patterns, and output formats unique to that field
  • Instruction Following: Train on diverse instruction-response pairs to improve the model's ability to follow complex multi-step instructions and produce structured outputs
  • Adapter Merging: After training, merge the LoRA adapter weights back into the base model with merge_and_unload() for inference without the PEFT overhead
  • Multi-task Training: Mix datasets from different tasks (summarization, classification, extraction) in a single fine-tuning run to create a versatile adapter

Pitfalls to Avoid

  • Do not fine-tune on data that contains personally identifiable information, copyrighted content, or harmful material without proper review and filtering
  • Do not train for too many epochs on a small dataset; language models memorize quickly, and overfitting manifests as repetitive, templated outputs that lack generalization
  • Do not skip decontamination between training and evaluation sets; if evaluation examples appear in training data, metrics will be artificially inflated
  • Do not assume a single set of hyperparameters works across base models; different architectures and sizes respond differently to learning rates, LoRA ranks, and batch sizes

Version History

  • acf2587 Current 2026-08-20 07:38

Same Skill Collection

crates/openfang-hands/bundled/browser/SKILL.md
crates/openfang-hands/bundled/clip/SKILL.md
crates/openfang-hands/bundled/collector/SKILL.md
crates/openfang-hands/bundled/infisical-sync/SKILL.md
crates/openfang-hands/bundled/lead/SKILL.md
crates/openfang-hands/bundled/predictor/SKILL.md
crates/openfang-hands/bundled/researcher/SKILL.md
crates/openfang-hands/bundled/trader/SKILL.md
crates/openfang-hands/bundled/twitter/SKILL.md
crates/openfang-skills/bundled/ansible/SKILL.md
crates/openfang-skills/bundled/api-tester/SKILL.md
crates/openfang-skills/bundled/aws/SKILL.md
crates/openfang-skills/bundled/azure/SKILL.md
crates/openfang-skills/bundled/ci-cd/SKILL.md
crates/openfang-skills/bundled/code-reviewer/SKILL.md
crates/openfang-skills/bundled/compliance/SKILL.md
crates/openfang-skills/bundled/confluence/SKILL.md
crates/openfang-skills/bundled/crypto-expert/SKILL.md
crates/openfang-skills/bundled/css-expert/SKILL.md
crates/openfang-skills/bundled/data-analyst/SKILL.md
crates/openfang-skills/bundled/data-pipeline/SKILL.md
crates/openfang-skills/bundled/docker/SKILL.md
crates/openfang-skills/bundled/elasticsearch/SKILL.md
crates/openfang-skills/bundled/email-writer/SKILL.md
crates/openfang-skills/bundled/figma-expert/SKILL.md
crates/openfang-skills/bundled/gcp/SKILL.md
crates/openfang-skills/bundled/git-expert/SKILL.md
crates/openfang-skills/bundled/github/SKILL.md
crates/openfang-skills/bundled/golang-expert/SKILL.md
crates/openfang-skills/bundled/graphql-expert/SKILL.md
crates/openfang-skills/bundled/helm/SKILL.md
crates/openfang-skills/bundled/interview-prep/SKILL.md
crates/openfang-skills/bundled/jira/SKILL.md
crates/openfang-skills/bundled/kubernetes/SKILL.md
crates/openfang-skills/bundled/linear-tools/SKILL.md
crates/openfang-skills/bundled/linux-networking/SKILL.md
crates/openfang-skills/bundled/ml-engineer/SKILL.md
crates/openfang-skills/bundled/mongodb/SKILL.md
crates/openfang-skills/bundled/nextjs-expert/SKILL.md
crates/openfang-skills/bundled/nginx/SKILL.md
crates/openfang-skills/bundled/notion/SKILL.md
crates/openfang-skills/bundled/oauth-expert/SKILL.md
crates/openfang-skills/bundled/openapi-expert/SKILL.md
crates/openfang-skills/bundled/postgres-expert/SKILL.md
crates/openfang-skills/bundled/presentation/SKILL.md
crates/openfang-skills/bundled/project-manager/SKILL.md
crates/openfang-skills/bundled/prometheus/SKILL.md
crates/openfang-skills/bundled/prompt-engineer/SKILL.md
crates/openfang-skills/bundled/python-expert/SKILL.md

Metadata

Files
0
Version
acf2587
Hash
3eb2461e
Indexed
2026-08-20 07:38

inicio - Wiki
Copyright © 2011-2026 iteam. Current version is 2.155.2. UTC+08:00, 2026-08-21 20:29
浙ICP备14020137号-1 $mapa de visitantes$