模型介绍
- 目标:通过ORPO技术对模型进行训练,以期达到以往指令微调加基于人类反馈的强化学习的效果
- 使用模型:LLaMA3-8B
- 使用数据集:mlabonne/orpo-dpo-mix-40k(共有数据44245条数据,仅使用了其中10000条数据)
- 使用显卡:RTX 4090,24G
- epoch:1
- per_device_train_batch_size=2
- gradient_accumulation_steps=4
模型使用
import transformers
import torch
model_id = "snowfly/llama3-8B-ORPO"
pipeline = transformers.pipeline("text-generation", model=model_id, model_kwargs={"torch_dtype": torch.bfloat16}, device_map="auto")
print(pipeline("Hey how are you doing today?"))
未完待续
- 使用的显卡显存不足,每个批次的数据量较少,训练中loss图上急剧震荡。后续再更多更大显存显卡上进行更大批处理数量上进行多轮训练
- 使用上述配置在全数据上训练3epoch需要72小时,实际实践使用其中随机选取的10000条数据训练1epoch
- Downloads last month
- 9
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.