菜单
Use when fine-tuning a large language model. Covers choosing full vs LoRA/QLoRA, dataset formatting, the transformers/PEFT/TRL stack, key hyperparameters, and evaluating fine-tunes without overfitting.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | llm-finetuning |
| description | Use when fine-tuning a large language model. Covers choosing full vs LoRA/QLoRA, dataset formatting, the transformers/PEFT/TRL stack, key hyperparameters, and evaluating fine-tunes without overfitting. |
Fine-tuning adapts a base LLM to a task or style. For almost all practitioners the right default is parameter-efficient fine-tuning (LoRA/QLoRA) — it trains <1% of weights, fits on a single GPU, and avoids catastrophic forgetting. Reach for full fine-tuning only with strong justification and budget.
Try prompting + few-shot + RAG first (see the rag-pipeline skill). Fine-tune when you need consistent format/style, lower latency/cost than long prompts, or to internalize a large example set.
| Method | Trains | VRAM | Use when |
|---|---|---|---|
| Full FT | 100% | Very high | Large data + budget, max quality |
| LoRA | ~0.1-1% | Moderate | Default for most tasks |
| QLoRA | ~0.1-1% | Low (4-bit) | Single consumer GPU |
Use the model's chat template; quality > quantity. A few thousand clean, diverse examples beat a noisy 100k dump.
def to_chat(example):
return {"messages": [
{"role": "system", "content": "You are a helpful support agent."},
{"role": "user", "content": example["question"]},
{"role": "assistant", "content": example["answer"]},
]}
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import torch
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map="auto")
peft_cfg = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05, bias="none",
task_type="CAUSAL_LM",
target_modules=["q_proj", "k_proj", "v_proj", "o_proj"])
trainer = SFTTrainer(
model=model, train_dataset=train_ds, eval_dataset=eval_ds, peft_config=peft_cfg,
args=SFTConfig(per_device_train_batch_size=2, gradient_accumulation_steps=8,
learning_rate=2e-4, num_train_epochs=2, warmup_ratio=0.03,
lr_scheduler_type="cosine", bf16=True, logging_steps=10,
eval_strategy="steps", eval_steps=100, output_dir="out"),
)
trainer.train()
r: 8-32 typical; raise for harder tasks.Merged or adapter weights + eval report, logged via experiment-tracking and deployed via model-serving.
Use when preparing raw data for modeling — handling missing values, duplicates, inconsistent types, outliers, and bad categorical values. Emphasizes fitting all imputation on train-only to avoid leakage.
Use when running ML experiments that need to be compared, reproduced, or shared. Covers MLflow/Weights & Biases logging, what to track, run organization, and model registry basics.
Use when starting on a new dataset, before modeling, or when asked to "explore", "profile", "understand", or "summarize" data. Covers structured EDA, distributions, correlations, target leakage checks, and visualization.
Use when creating, encoding, scaling, or selecting features for ML models. Covers categorical encoding, numeric transforms, datetime/text/aggregation features, and leakage-safe target encoding.
Use when optimizing model hyperparameters. Covers search strategy (random vs Bayesian/Optuna), leakage-safe tuning inside CV, search-space design, early stopping, and budget management.
Use when the target is rare (fraud, churn, disease, anomalies). Covers correct metrics, resampling (SMOTE/undersampling), class weights, threshold tuning, and avoiding the accuracy trap and resampling leakage.