JobCannon
全部技能

LLM Fine-Tuning

Customizing large language models for domain-specific applications

⬢ 層級 3技術
+$40k-
薪資影響
8 個月
學習時長
困難
難度
12
職業
一覽

LLM fine-tuning adapts foundation models (GPT, Llama, Mistral, Claude) to domain-specific tasks with parameter-efficient methods (LoRA/QLoRA) or full training. Career path: Practitioner (OpenAI API, basic data prep, $140-170k) → Specialist (LoRA/RLHF, Hugging Face ecosystem, $160-210k) → Expert (distributed training, custom objectives, $200-280k). Salaries top-tier: USA $130-280k, UK £75-160k, EU €85-175k. Techniques: LoRA (low-rank adaptation, 10% compute of full tune), QLoRA (quantized, fit on single GPU), RLHF/DPO (alignment), evaluation frameworks. When to fine-tune: if domain-specific performance >>general model, or budget allows; otherwise RAG or prompt engineering may suffice.

什麼是 LLM Fine-Tuning

LLM fine-tuning adapts pre-trained language models to specific domains, tasks, or styles. Techniques range from full fine-tuning to parameter-efficient methods (LoRA, QLoRA) that require minimal compute. Fine-tuning enables creating specialized models that outperform general-purpose LLMs for specific use cases. Understanding when to fine-tune vs use prompt engineering or RAG, and how to prepare training data, is a critical skill for AI engineers building production AI applications.

🔧 工具與生態
HuggingFace TransformersPEFTTRLAxolotlUnslothMLflowWeights & BiasesModalReplicateOpenAI fine-tuning API

💰 各地區薪資

地區初級中級高級
USA$140,000$210,000$280,000
UK£75,000£117,500£160,000
EU€85,000€130,000€175,000
CANADACA$145,000CA$217,500CA$290,000

❓ 常見問題

Full fine-tune vs LoRA, which should I use?
LoRA: 1-2% of parameters, 90% faster, fits single GPU, 90% of quality gain. Full fine-tune: all parameters, slower, needs multi-GPU, slightly higher quality, overfitting risk on small datasets. Start with LoRA for exploration; full fine-tune only if you have >100k labeled examples and compute budget.
When should I fine-tune vs use RAG vs better prompting?
Prompting: < 1 week, free, baseline. RAG: 1-2 weeks, retrieval-based, no training data needed, good for knowledge updates. Fine-tuning: 2-8 weeks, requires labeled data, permanent model changes, best when general model fundamentally misses your domain (medical, legal, code).
How much training data do I need for LoRA?
LoRA: 500-5000 examples (domain-specific task). Full fine-tune: 5000-100k+. Quality > quantity: 500 excellent examples beat 10k mediocre ones. BLEU, ROUGE, or task-specific metrics guide sufficiency. Start with 10% of your budget to validate data quality.
What's RLHF and DPO, and when do I use them?
RLHF (Reinforcement Learning from Human Feedback): align model to human preferences via reward model, computationally expensive, industry standard (ChatGPT). DPO (Direct Preference Optimization): simpler, no reward model, same alignment quality, emerging 2024-2026. Use DPO for most cases unless you have production RLHF infrastructure.
How do I evaluate a fine-tuned model?
Automated: BLEU/ROUGE (translation/summarization), METEOR, exact match (QA). Human: n=100 blind test vs baseline. Business: latency, cost, throughput. Always test on held-out set (10-20% of data) to catch overfitting. Compare model outputs qualitatively.
What are GPU cost implications for fine-tuning?
LoRA on single GPU (H100): $2-8/hour. Full fine-tune on multi-GPU: $50-300+/hour. 1-week project: $2k-10k total. OpenAI API (per 1M tokens): $1.50-$30 depending on model. Calculate ROI: if fine-tuned model saves 10% on inference costs or increases quality by 20%, payoff within months.
Which base model should I fine-tune, 7B, 13B, 70B, or use an API?
7B (Mistral, Llama-2): fastest, cheapest, edge deployment, good for latency-critical tasks. 13B-70B: better quality, slower. 70B: near-GPT4 but expensive to run. API fine-tuning (OpenAI, Anthropic): zero ops, blackbox, most expensive. For startups: start 7B LoRA locally, graduate to 13B, switch to API only if custom tuning impossible.

不確定這項技能適不適合你？

做一下職業匹配，我們會推薦合適的方向。

找到最適合我的技能 →

找到理想的職業方向

在 2,521 個職業中進行技能匹配。免費。

免費做職業匹配 →