Init run 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132

6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132/README.md

@@ -0,0 +1,26 @@
+---
+library_name: nemo
+tags:
+- sortformer
+- speaker-diarization
+- automatic-speech-recognition
+---
+
+# 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
+
+Sortformer fine-tuning run, hosted under [arda-argmax/sortformer-finetuned](https://huggingface.co/arda-argmax/sortformer-finetuned).
+
+## Best checkpoint
+
+- **monitor**: `val_der` (min)
+
+Files:
+- `best.nemo` — full NeMo checkpoint (config + weights), restorable via `SortformerEncLabelModel.restore_from('best.nemo')`
+- `best.json` — machine-readable metadata for the best checkpoint
+- `checkpoints/step_*_epoch_*_*.nemo` — archived copies of every checkpoint that became the best during training
+- `config.yaml` — Hydra config the run was launched with
+
+## Run info
+
+- started: `2026-06-04T05:05:06Z`
+- live dashboard: [arda-argmax/sortformer-trackio-v2](arda-argmax/sortformer-trackio-v2)

6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132/config.yaml

@@ -0,0 +1,114 @@
+seed: 42
+force_exit_on_completion: true
+model:
+  init_checkpoint: arda-argmax/diar_streaming_sortformer_6spk-init
+  num_spks: 6
+data:
+  cuts_path: null
+  dataset_hf_repo: null
+  split: null
+  cuts_filename: null
+  preprocess: true
+  audio_root: null
+  session_len_sec: 120
+  sample_rate: 16000
+  batch_size: 12
+  num_workers: 4
+  shuffle: true
+  prepared:
+    enabled: true
+    uri: gs://fastmss-synth/sortformer/fastmss_synth/v0.6_rttm_segment_opus96k
+    local_dir: /root/fastmss_prepared_shar
+    root_dir: null
+    train_dir: null
+    val_dir: null
+    train_path: train
+    val_path: val
+    sync: true
+    gcloud_bin: gcloud
+    max_train_cuts: null
+    max_val_cuts: null
+    num_train_samples: 2996000
+  streaming:
+    enabled: false
+    repo: arda-argmax/simchoir-parquet
+    subset: v0.1
+    label_source: legacy
+    cache_materialization: stream
+    download_dir: /tmp/sortformer_hf_parquet_chunks
+    materialization_workers: 1
+    subset_prefix: null
+    train_subsets: null
+    val_subsets: null
+    train_split: train
+    val_split: val
+    train_shuffle_buffer: 0
+    sampler_shuffle_buffer: 1000
+    num_train_samples: 800
+    train_cache_size: -1
+    cache_storage: disk
+    cache_dir: /tmp/sortformer_streaming_cache
+    val_cache_size: -1
+    val_replicate_across_ranks: false
+    bucketing:
+      enabled: false
+      max_duration_sec: null
+      num_buckets: 10
+      buffer_size: 256
+      quadratic_duration: 30.0
+      duration_bins: []
+      bin_estimate_samples: 10000
+optim:
+  learning_rate: 5.0e-05
+  weight_decay: 0.0
+  warmup_steps: 50
+  min_lr: 1.0e-06
+trainer:
+  accelerator: gpu
+  devices: auto
+  precision: bf16-mixed
+  fallback: false
+  num_steps: -1
+  num_epochs: 3
+  val_check_interval: 0.05
+  check_val_every_n_epoch: 1
+  log_every_n_steps: 1
+  limit_val_batches: 1.0
+  num_sanity_val_steps: 0
+  accumulate_grad_batches: 1
+  strategy: auto
+regularization:
+  disable_all_dropout: false
+  causal_attn_rate: null
+  scores_add_rnd: null
+output:
+  final_checkpoint_save_path: /app/training_runs/1780546132/final.nemo
+  log_dir: /app/training_runs/1780546132/logs
+  run_name: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
+tracking:
+  enabled: true
+  project: sortformer-robustness
+  run_name: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
+  group: null
+  space_id: null
+  auto_log_gpu: true
+  gpu_log_interval: 10.0
+  console_log:
+    enabled: false
+    max_chars: 200000
+    every_n_train_steps: 0
+val_der:
+  enabled: true
+  collar: 0.0
+  skip_overlap: false
+  threshold: 0.5
+  max_samples: null
+checkpointing:
+  enabled: true
+  repo: arda-argmax/sortformer-finetuned
+  run_subdir: 6spk-v0.6-fastmss-prepared-shar-bs12-a100-80gb-2gpu-fullset-v4-1780546132
+  monitor: val_der
+  mode: min
+  private: true
+  local_dir: null
+  flush_timeout_sec: 1800