视频总结

23年发布的模型在一些材料中归位指令微调模型，后面逐渐升级应该已经是train的模型了

技术报告总结

InternLM2 Technical Report

评测与特点

• 6 dimensions and 30 benchmarks, long-context modeling, and open-ended
  subjective evaluations
• 长文本，目前已经表明硬train是可以实现大海捞针， long-term dependencies, initially trained on 4k tokens before advancing to 32k tokens in pre-training and fine-tuning stages

数据

• diverse data types including text, code, and long-context data. web pages, papers, patents, and books
• initially trained on 4k tokens before advancing to 32k tokens in pre-training
• processing steps including rule-based filtering, data deduplication, safety filtering, and quality filtering

预训练

对齐

• Supervised Fine-Tuning (SFT) 和一种新的强化学习 COOL RLHF

尝试