# rasr-parakeet-v2: continued finetune of v1 on real US ATC (TartanAviation).
#
# v1 was strong on European ATC but weak on US ATC; its only real-audio anchor
# was European. v2 adds about 29 h of real US tower audio and closes most of
# that gap: liveATC WER 0.311 to 0.251, European WER held flat.
#
# This is a continued finetune from v1's weights (model.ref is the v1 .nemo),
# not a train from base. 10,000 steps at lr 5e-5.
#
# The US training manifest is built from the published TartanAviation datasets
# (twangodev/tartanaviation-atc-adsb-utterances audio plus ...-atc-labels pseudo
# labels): keep clips with inter-model agreement >= 0.9 and non-empty text, join
# audio to labels by row order, and exclude the held-out evaluation dates. About
# 35,240 clips / 28.8 h. The EU train splits are light rehearsal to limit forgetting.

defaults: [base, rtx6kpro/hw]

name: rasr-parakeet-v2

model:
  scheme: parakeet
  ref: rasr-parakeet-v1.nemo        # continued finetune from v1 (hf:twangodev/rasr-parakeet-v1)

data:
  train:
    - dataset: local:data/tartan/us_train.jsonl   # TartanAviation, agreement >= 0.9, held-out dates excluded
      weight: 1.0
    - dataset: hf:jlvdoorn/atco2-asr:train
      weight: 1.0                    # European rehearsal
    - dataset: hf:jlvdoorn/atco2-asr-atcosim:train
      weight: 1.0
  validation:
    - dataset: hf:jlvdoorn/atco2-asr:validation

optimizer:
  lr: 5.0e-5

scheduler:
  warmup_steps: 300
  min_lr: 1.0e-6

trainer:
  max_steps: 10000
  val_check_interval: 2000
  num_workers: 4

output:
  dir: ckpt/${name}