Issues

• Log the effective LR (and MLP settings) at training startup

  enhancement
  Oaklight/krino#125 · Filter by author Oaklight opened on Oct 14, 2026

• sampling_temperature is the reciprocal of the mT5 alpha it cites, so the temperature sweep tested the wrong regime

  area:dataarea:trainingbug
  Oaklight/krino#121 · Filter by author Oaklight opened on Oct 13, 2026
  2

• Synthetic corpus: near-duplicate states leak across the train/test split

  area:databug
  Oaklight/krino#120 · Filter by author Oaklight opened on Oct 13, 2026

• Epic: Data pipeline — synthetic generation, augmentation, and mixing

  area:dataepic
  Oaklight/krino#106 · Filter by author Oaklight opened on Sep 27, 2026
  1

• Noul-to-Choice augmentation: convert binary questions to multi-option with LLM/Jev-generated alternatives

  area:dataarea:modelarea:researchenhancement
  Oaklight/krino#104 · Filter by author Oaklight opened on Sep 27, 2026
  6

• Data mixing: task/source ratio optimization for multi-task training

  area:dataarea:trainingenhancementresearch
  Oaklight/krino#97 · Filter by author Oaklight opened on Sep 26, 2026
  3

• Multi-task optimization: per-type LR, gradient balancing, adaptive loss weighting

  area:researcharea:trainingenhancementresearch
  Oaklight/krino#95 · Filter by author Oaklight opened on Sep 26, 2026
  4

• Experiment: MLP projector + per-component LR (frozen backbone done, LoRA pending)

  area:researcharea:trainingenhancementresearch
  Oaklight/krino#93 · Filter by author Oaklight opened on Sep 26, 2026
  4

• CLM-8B: open-source System One model with contrastive training (Stanford/NVIDIA)

  area:researchdocumentationresearch
  Oaklight/krino#76 · Filter by author Oaklight opened on Sep 25, 2026
  1

• Literature review: LLM routing models and transferable calibration strategies

  area:researchdocumentationresearch
  Oaklight/krino#74 · Filter by author Oaklight opened on Sep 24, 2026

• Epic: Training pipeline — supervised → multi-task → calibration → PEFT

  area:trainingepic
  Oaklight/krino#73 · Filter by author Oaklight opened on Sep 24, 2026
  2

• Web: interactive benchmark dashboard + live inference playground (GitHub Pages)

  documentation
  Oaklight/krino#63 · Filter by author Oaklight opened on Sep 23, 2026

• Future: hybrid encoder-decoder decision model

  architecturearea:modelarea:researchdeferredenhancement
  Oaklight/krino#60 · Filter by author Oaklight opened on Sep 22, 2026

• Future: compact decision model (<10M params)

  architecturearea:modelarea:researchdeferredenhancement
  Oaklight/krino#59 · Filter by author Oaklight opened on Sep 22, 2026
  1

• Step 3: Calibration training — losses + RLCD-style RL

  area:modelarea:trainingenhancement
  Oaklight/krino#23 · Filter by author Oaklight opened on Sep 19, 2026
  2

• Epic: Decision model replication

  area:modelarea:researchepic
  Oaklight/krino#19 · Filter by author Oaklight opened on Sep 19, 2026
  3