dehanns/whisper-small-sequential-cs-adapter

1{2  "args": {3    "config": "/root/sequential_adapter_strategy/configs/sequential_cs.yaml",4    "base_model": "openai/whisper-small",5    "si_adapter": "dehanns/whisper-small-lora-sinhala-32",6    "si_adapter_revision": "epoch-5",7    "expected_si_rank": 32,8    "init_cs_adapter": null,9    "train_source": "/experiment/splits/rich_codeswitch_v2_train.csv",10    "eval_source": "/experiment/splits/rich_codeswitch_v2_validation.csv",11    "si_eval_source": "none",12    "splits_dir": "/experiment/splits",13    "metadata_csv": null,14    "output_dir": "/experiment/adapters/lora_cs_on_frozen_si",15    "cache_dir": "/experiment/hf-cache",16    "cs_rank": 16,17    "cs_lora_alpha": 32,18    "cs_lora_dropout": 0.05,19    "target_modules": "q_proj,v_proj",20    "epochs": 6,21    "learning_rate": 0.0001,22    "batch_size": 16,23    "grad_accum": 1,24    "eval_batch_size": 8,25    "warmup_steps": 50,26    "max_train_samples": null,27    "max_eval_samples": null,28    "max_si_eval_samples": 100,29    "num_proc": null,30    "lazy_loading": true,31    "dataloader_num_workers": 4,32    "monitor_metric": "cs_wer",33    "language": "sinhala",34    "task": "transcribe",35    "seed": 42,36    "fp16": true37  },38  "strategy": {39    "strategy": "frozen_sinhala_plus_trainable_cs_residual",40    "checkpoint_kind": "report",41    "base_model": "openai/whisper-small",42    "required_sinhala_adapter": "dehanns/whisper-small-lora-sinhala-32",43    "required_sinhala_revision": "epoch-5",44    "expected_sinhala_rank": 32,45    "required_inference_adapters": [46      "si",47      "cs"48    ],49    "required_inference_weights": {50      "si": 1.0,51      "cs": 1.052    },53    "cs_initialization": "fresh",54    "cs_adapter": {55      "rank": 16,56      "lora_alpha": 32,57      "lora_dropout": 0.05,58      "target_modules": [59        "q_proj",60        "v_proj"61      ]62    },63    "note": "This CS residual is not intended for standalone use."64  },65  "freeze_report_before": {66    "active_adapters": [67      "si",68      "default"69    ],70    "trainable_parameter_tensors": 144,71    "trainable_parameters": 1769472,72    "frozen_si_parameter_tensors": 144,73    "disabled_si_dropout_modules": 72,74    "si_rank": 32,75    "cs_rank": 1676  },77  "freeze_report_after": {78    "active_adapters": [79      "si",80      "default"81    ],82    "trainable_parameter_tensors": 144,83    "trainable_parameters": 1769472,84    "frozen_si_parameter_tensors": 144,85    "disabled_si_dropout_modules": 7286  },87  "sinhala_fingerprint_before": "8601dd13f7d54149c2d6ca1e1d413d2b3ded13b2db272f65477d5ca1922a036a",88  "sinhala_fingerprint_after": "8601dd13f7d54149c2d6ca1e1d413d2b3ded13b2db272f65477d5ca1922a036a",89  "sinhala_unchanged": true,90  "train_metrics": {91    "train_runtime": 1059.2354,92    "train_samples_per_second": 14.076,93    "train_steps_per_second": 0.884,94    "total_flos": 4.4167821336576e+18,95    "train_loss": 0.1281622916332677,96    "epoch": 6.097  },98  "validation_history": [99    {100      "epoch": 1.0,101      "step": 156,102      "eval_loss": 0.18984781205654144,103      "codeswitch": {104        "wer": 19.453924914675767,105        "cer": 6.7952249770431585,106        "n_samples": 61107      },108      "sinhala_retention": null109    },110    {111      "epoch": 2.0,112      "step": 312,113      "eval_loss": 0.15295149385929108,114      "codeswitch": {115        "wer": 15.813424345847555,116        "cer": 5.215794306703398,117        "n_samples": 61118      },119      "sinhala_retention": null120    },121    {122      "epoch": 3.0,123      "step": 468,124      "eval_loss": 0.1432490348815918,125      "codeswitch": {126        "wer": 15.244596131968146,127        "cer": 5.4729109274563825,128        "n_samples": 61129      },130      "sinhala_retention": null131    },132    {133      "epoch": 4.0,134      "step": 624,135      "eval_loss": 0.13980381190776825,136      "codeswitch": {137        "wer": 14.106939704209328,138        "cer": 5.252525252525253,139        "n_samples": 61140      },141      "sinhala_retention": null142    },143    {144      "epoch": 5.0,145      "step": 780,146      "eval_loss": 0.1419447809457779,147      "codeswitch": {148        "wer": 16.1547212741752,149        "cer": 7.309458218549128,150        "n_samples": 61151      },152      "sinhala_retention": null153    },154    {155      "epoch": 6.0,156      "step": 936,157      "eval_loss": 0.14258421957492828,158      "codeswitch": {159        "wer": 15.017064846416384,160        "cer": 5.766758494031221,161        "n_samples": 61162      },163      "sinhala_retention": null164    }165  ],166  "best_epoch": 4.0,167  "best_metric": {168    "cs_wer": 14.106939704209328169  }170}