低成本微调垂直领域专属ChatGPT
1.大模型微调常见方法和步骤
关于FinetuningPipeline
- Pre-training——Generic Data
- Continual Pre-training——Domain-Specific Data
- Supervised Finetuning——Instructional Promt Data
- RLHF
LMFlow:
- Continual Pre-training
- Task Tuning
- INstruction Tuning
- Parameter-Efficient Tuning
- Aligment Tuning
- Large Model Inference
2.低成本微调大模型的全流程框架LMFlow
容易扩展支持海量模型、支持任意定制化数据上的快速训练、轻量级、低成本。
3.一种全自动可扩展的模型评测基准LMFlow Benchmark
Human Evaluation
Rouge_L
Accuracy
GPT-4 as a judge
Negative Log Likelihood(ours)