I use in both cases q4_K_M

  • VertexMachine@alien.topB
    link
    fedilink
    English
    arrow-up
    1
    ·
    11 months ago

    It’s a source. But rarely synthetic benchmarks give you the whole picture. Plus those test sets are in the public, so there is some incentive for some people to game the system (and even without that those data sets most likely are already in the training data).

    • TobyWonKenobi@alien.topB
      link
      fedilink
      English
      arrow-up
      1
      ·
      11 months ago

      I’ve had the same experience. Are you using GGUF? I do, and I’ve heard that Yi may suffer from GGUF. So EXL2 might be better… I need to try it and see.