ChatQA: Building GPT-4 Level Conversational QA Models
Meta info.
- Authors: ihan Liu, Wei Ping, Rajarshi Roy, Peng Xu, Mohammad Shoeybi, Bryan Catanzaro
- Paper: https://arxiv.org/abs/2401.10225
- Affiliation: NVIDIA
TL; DR
LLM zero-shotμμ λνκΌ΄ QA μ±λ₯μ ν¬κ² κ°μ ν μ μλ 2-stage instruction tuning λ°©λ² μ μ.
Suggestions
- stage 1: multi-turn λνλ°μ΄ν°λ‘ SFT
- stage 2: λ§₯λ½μ΄ μ£Όμ΄μ§λ QA λ²€μΉλ§ν¬ λ°μ΄ν°λ‘ instruction tuning
- retrieval for multi-turn QA: λνκ° κΈΈμ΄μ§ κ²½μ°, μ§μ λ°νμ λν μ΄λ ₯μ μΈμ½λ©ν΄μ κ΄λ ¨ λν λΆλΆμ μ°Ύμμ¨λ€κ³ . (pic2)