Issues

• Huge discrepancy between training PRED and generation inference results on Llama-3-8B.Bad performance for generation modes.

  microsoft/KBLaM#96 · Filter by author linke626 opened on Feb 6, 2026

• About KB_size during training

  microsoft/KBLaM#94 · Filter by author gwyong opened on Oct 7, 2025

• Available pre-trained linear adapter's weights

  microsoft/KBLaM#93 · Filter by author gwyong opened on Oct 6, 2025

• Issue with Embedding shape

  microsoft/KBLaM#92 · Filter by author pTidke opened on Oct 3, 2025

• Future model support (Qwen3 dens/moe)

  microsoft/KBLaM#90 · Filter by author Chen-DaYi opened on Aug 29, 2025
  2

• Working with newer Transformer versions

  microsoft/KBLaM#89 · Filter by author vpgits opened on Aug 27, 2025
  6

• Guidance for Integrating KBLaM with Food/Nutrition Knowledge Base

  microsoft/KBLaM#82 · Filter by author maksvavken opened on Jul 29, 2025
  7

• Where is the sep_query_head stored?

  microsoft/KBLaM#81 · Filter by author AfiaI14 opened on Jul 23, 2025
  1

• ImportError: cannot import name 'LlamaDynamicNTKScalingRotaryEmbedding' from 'transformers.models.llama.modeling_llama'

  microsoft/KBLaM#80 · Filter by author shiwanghua opened on Jul 23, 2025
  1

• In experiments/train.py，training labels problem

  microsoft/KBLaM#79 · Filter by author Jean-ZW opened on Jul 21, 2025
  2

• Has anyone had a good trained model

  microsoft/KBLaM#75 · Filter by author retoriax opened on Jul 8, 2025
  8

• Bug: encoder.pt and config not loaded correct during resume

  microsoft/KBLaM#72 · Filter by author retoriax opened on Jun 26, 2025
  1

• Readme train.py command doesn't create saved output once training is complete

  microsoft/KBLaM#71 · Filter by author grctest opened on Jun 24, 2025
  8

• eval kb generation mode does not work well

  microsoft/KBLaM#70 · Filter by author Chloe-mxxxxc opened on Jun 23, 2025

• Questions about the Rouge metrics

  microsoft/KBLaM#66 · Filter by author yujia-fu opened on Jun 11, 2025
  10

• Bad performance for three generation modes.

  microsoft/KBLaM#64 · Filter by author shiwanghua opened on May 28, 2025
  20

• Question about train.py

  microsoft/KBLaM#62 · Filter by author laker123456 opened on May 26, 2025
  5

• Question about eval parameter.

  microsoft/KBLaM#61 · Filter by author shiwanghua opened on May 22, 2025
  5

• Training Effects

  microsoft/KBLaM#60 · Filter by author laker123456 opened on May 21, 2025

• There is a formula missing an 'l' in the paper

  microsoft/KBLaM#57 · Filter by author shiwanghua opened on May 19, 2025
  8

• Bug in train_utils.py

  microsoft/KBLaM#56 · Filter by author shiwanghua opened on May 16, 2025
  11

• Has anyone trained with Phi models before? How was the effectiveness?

  microsoft/KBLaM#55 · Filter by author laker123456 opened on May 16, 2025
  1

• Question about dataset

  microsoft/KBLaM#52 · Filter by author shiwanghua opened on May 16, 2025
  2

• Questions concerning train.py and its parameters

  microsoft/KBLaM#51 · Filter by author ThomasHoppe opened on May 1, 2025
  1

• Generating KB embeddings does not work if AZURE OPENAI API KEY is set

  microsoft/KBLaM#49 · Filter by author ThomasHoppe opened on Apr 24, 2025