Fine-Tuning微调大模型

When to Use

User wants to fine-tune a language model, evaluate if fine-tuning is worth it, or debug training issues.

Quick Reference

Topic	File
Provider comparison & pricing	INLINECODE0
Data preparation & validation

data-prep.md | | Training configuration | training.md | | Evaluation & debugging | evaluation.md | | Cost estimation & ROI | costs.md | | Compliance & security | compliance.md |

Core Capabilities

1. 1. Decide fit — Analyze if fine-tuning beats prompting for the use case
2. Prepare data — Convert raw data to JSONL, deduplicate, validate format
3. Select provider — Compare OpenAI, Anthropic (Bedrock), Google, open source based on constraints
4. Estimate costs — Calculate training cost, inference savings, break-even point
5. Configure training — Set hyperparameters (learning rate, epochs, LoRA rank)
6. Run evaluation — Compare fine-tuned vs base model on task-specific metrics
7. Debug failures — Diagnose loss curves, overfitting, catastrophic forgetting
8. Handle compliance — Scan for PII, configure on-premise training, generate audit logs

Decision Checklist

Before recommending fine-tuning, ask:

• - [ ] What's the failure mode with prompting? (format, style, knowledge, cost)
• [ ] How many training examples available? (minimum 50-100)
• [ ] Expected inference volume? (affects ROI calculation)
• [ ] Privacy constraints? (determines provider options)
• [ ] Budget for training + ongoing inference?

Fine-Tune vs Prompt Decision

Signal	Recommendation
Format/style inconsistency	Fine-tune ✓
Missing domain knowledge

RAG first, then fine-tune if needed | | High inference volume (>100K/mo) | Fine-tune for cost savings | | Requirements change frequently | Stick with prompting | | <50 quality examples | Prompting + few-shot |

Critical Rules

• - Data quality > quantity — 100 great examples beat 1000 noisy ones
• LoRA first — Never jump to full fine-tuning; LoRA is 10-100x cheaper
• Hold out eval set — Always 80/10/10 split; never peek at test data
• Same precision — Train and serve at identical precision (4-bit, 16-bit)
• Baseline first — Run eval on base model before training to measure actual improvement
• Expect iteration — First attempt rarely optimal; plan for 2-3 cycles

Common Pitfalls

Mistake	Fix
Training on inconsistent data	Manual review of 100+ samples before training
Learning rate too high

Start with 2e-4 for SFT, 5e-6 for RLHF | | Expecting new knowledge | Fine-tuning adjusts behavior, not knowledge — use RAG | | No baseline comparison | Always test base model on same eval set | | Ignoring forgetting | Mix 20% general data to preserve capabilities |