# Second LoRA round: answerability for RAG (can this document answer this question?), on top of hayriyigit/decider-4b-tr.
#   scripts/answerability_run.sh   (builds data/answerability_mix, scores the starting model, then this)
workdir: ./answerability_run

model:
  repo: hayriyigit/decider-4b-tr                # private: HF_TOKEN in .env
  revision: df9aacb9f4c9e4e64178176502f51ae4db271f26

data:
  train: data/answerability_mix/train-00000-of-00001.parquet   # scripts/build_answerability_mix.py
  test: data/answerability_mix/test-00000-of-00001.parquet
  rows:                                         # everything the mix holds; replay rows train toward the starting model
    wiki2_tr: 24000                             # same attribute, other entity (the judge set's entity errors)
    hotpot_tr: 30000                            # all hops / a missing hop / on-topic distractors only
    musique_tr: 10000                           # all hops (judge-checked) / a missing hop
    synth_tr: 30000                             # Turkish pairs: institution, entity, time, place/facet, attribute, partial
    squad_tr: 12000
    replay_bev_tr: 4000
    replay_bev_tr_en: 3000
  holdout_rows: 1500
  max_state_tokens: 4096
  seed: 0

lora:
  r: 64
  alpha: 128

train:
  epochs: 1
  lr: 1.0e-4
  warmup: 0.05
  tokens_per_step: 65536
  micro_tokens: 16384

checkpoint:
  every_minutes: 30

limits:
  max_hours: 6
  reserve_minutes: 20

export:
  push_to_hub: {repo: hayriyigit/decider-4b-tr-rag, private: true}