All pull requests

Pull requests

• Compute corresponding BGE-M3 inference scores with tiled MaxSim

  #1637·htripathi12 opened Nov 13, 2026·

  ·
  1

• Align M3 ColBERT vectors and masks across variable-length sub-batches

  #1636·Afloat16 opened Nov 12, 2026·

  ·

• Apply retrieval and reranking depth after candidate filtering

  #1635·Afloat16 opened Nov 12, 2026·

  ·

• Include missing retrieval queries in evaluation averages

  #1633·Afloat16 opened Nov 12, 2026·

  ·

• Preserve retrieval candidates when evaluating rerankers

  #1634·Afloat16 opened Nov 12, 2026·

  ·

• Handle empty evaluation reranking candidates

  #1632·Afloat16 opened Nov 12, 2026·

  ·

• Avoid cross-query score matrices for local-negative training

  #1631·Afloat16 opened Nov 12, 2026·

  ·

• Train all requested layerwise reranker heads

  #1630·Afloat16 opened Nov 12, 2026·

  ·

• Preserve reranker query groups in partial training batches

  #1629·Afloat16 opened Nov 12, 2026·

  ·

• Keep unprompted same-dataset training text unchanged

  #1628·Afloat16 opened Nov 12, 2026·

  ·

• Load training files in a stable order across ranks

  #1627·Afloat16 opened Nov 12, 2026·

  ·

• Pull towards Junji accross the Flags model

  #1626·sheikhsajid69 opened Nov 12, 2026·

  ·

• Optimize BGE-M3 compute_score with pairwise scoring (#1624)

  #1625·Rajan-vish opened Nov 11, 2026·

  ·
  1

• fix: propagate inference worker failures to callers

  #1623·remote-controlled-man opened Nov 7, 2026·

  ·
  1

• fix: preserve dense pooling in encoder checkpoints

  #1622·emme1t opened Nov 1, 2026·

  ·
  1

• fix: load auto models from paths with trailing separators

  #1621·emme1t opened Nov 1, 2026·

  ·
  1

• fix: require explicit opt-in for embedder remote code

  #1620·original4422 opened Oct 31, 2026·

  ·
  1

• fix: preserve layerwise scores across reranker worker batches

  #1619·linjiaye929 opened Oct 31, 2026·

  ·

• fix(m3): score a single pair on multiple devices

  #1616·michaelkillgta opened Oct 12, 2026·

  ·
  1

• Fix distillation with integer teacher scores

  #1613·zhechen-cs opened Oct 11, 2026·

  ·
  1

• Pool M3 training sparse weights without token-vocabulary expansion

  #1612·Afloat16 opened Oct 7, 2026·

  ·

• Exclude passage padding from M3 training MaxSim

  #1610·Afloat16 opened Oct 2, 2026·

  ·

• Keep per-query negative groups intact in M3 training

  #1609·Afloat16 opened Oct 2, 2026·

  ·

• fix(inference): only retry batch-size probing on genuine OOM errors

  #1608·rtmalikian opened Sep 30, 2026·

  ·
  1

• fix(eval): rank a tied document by id before insertion order

  #1606·SashaMIT opened Sep 29, 2026·

  ·