{
  "contains_mixed_model_types": false,
  "metric_recommendations": {
    "cloze_en_contains": {
      "checkpoint_name": "step_9000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_9000.pt",
      "direction": "max",
      "value": 0.02
    },
    "cloze_en_exact": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.0
    },
    "cloze_it_contains": {
      "checkpoint_name": "step_8000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_8000.pt",
      "direction": "max",
      "value": 0.14
    },
    "cloze_it_exact": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.0
    },
    "distinct_1": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "max",
      "value": 0.24277456647398843
    },
    "distinct_2": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "max",
      "value": 0.49899799599198397
    },
    "language_consistency_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.95
    },
    "language_consistency_it": {
      "checkpoint_name": "step_8000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_8000.pt",
      "direction": "max",
      "value": 0.85
    },
    "language_switch_rate_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 0.0
    },
    "language_switch_rate_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 0.0
    },
    "loop_rate": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 0.4
    },
    "ppl_en": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 147.3916153866132
    },
    "ppl_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 56.145236622362084
    },
    "ppl_mixed": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 186.23824092873335
    },
    "repeated_4gram_rate": {
      "checkpoint_name": "step_8000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_8000.pt",
      "direction": "min",
      "value": 0.85
    },
    "source_loss_books_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 4.718399774460566
    },
    "source_loss_books_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 4.800248645600819
    },
    "source_loss_code": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_10000.pt",
      "direction": "min",
      "value": 7.710747273763021
    },
    "source_loss_web_en": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_10000.pt",
      "direction": "min",
      "value": 5.753730773925781
    },
    "source_loss_web_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 6.1019945646587175
    },
    "source_loss_wiki_en": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 4.047422235662287
    },
    "source_loss_wiki_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 3.762100830078125
    },
    "val_loss_en": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 4.993093094735776
    },
    "val_loss_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 4.0279418445024335
    },
    "val_loss_mixed": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "direction": "min",
      "value": 5.227026719313401
    }
  },
  "recommended_checkpoint": {
    "checkpoint_name": "step_11000",
    "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
    "direction": "min",
    "value": 5.227026719313401
  },
  "recommended_metric": "val_loss_mixed",
  "rows": [
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_7000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 7000,
      "cloze_en_contains": 0.0,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.08,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.21282798833819241,
      "distinct_2": 0.43377148634984836,
      "generation_pass_rate": null,
      "generation_pass_rate_regression_vs_previous": false,
      "generation_passed_prompts": 0,
      "generation_scored_prompts": 0,
      "generation_total_prompts": 40,
      "language_consistency_en": 0.95,
      "language_consistency_it": 0.775,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.6,
      "model_type": "pretrained",
      "ppl_en": 157.69045491120713,
      "ppl_it": 56.14792618913685,
      "ppl_mixed": 186.36645615831696,
      "repeated_4gram_rate": 0.925,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 4.718399774460566,
      "source_loss_books_it": 4.809140159970238,
      "source_loss_code": 8.382174173990885,
      "source_loss_web_en": 6.0977221037212175,
      "source_loss_web_it": 6.1019945646587175,
      "source_loss_wiki_en": 4.11204043301669,
      "source_loss_wiki_it": 3.92701171875,
      "val_loss_en": 5.0606339652583285,
      "val_loss_it": 4.027989747094327,
      "val_loss_mixed": 5.22771492982522,
      "validation_loss_regression_vs_previous": false
    },
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_8000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_8000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 8000,
      "cloze_en_contains": 0.0,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.14,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.20623281393217233,
      "distinct_2": 0.4262607040913416,
      "generation_pass_rate": null,
      "generation_pass_rate_regression_vs_previous": false,
      "generation_passed_prompts": 0,
      "generation_scored_prompts": 0,
      "generation_total_prompts": 40,
      "language_consistency_en": 0.95,
      "language_consistency_it": 0.85,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.525,
      "model_type": "pretrained",
      "ppl_en": 153.74673622063136,
      "ppl_it": 65.00039768395027,
      "ppl_mixed": 200.537199596494,
      "repeated_4gram_rate": 0.85,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 5.201952979678199,
      "source_loss_books_it": 5.141731988816034,
      "source_loss_code": 7.883303833007813,
      "source_loss_web_en": 5.875253376207854,
      "source_loss_web_it": 6.312699970446135,
      "source_loss_wiki_en": 4.081960851495916,
      "source_loss_wiki_it": 4.03841064453125,
      "val_loss_en": 5.035306678628022,
      "val_loss_it": 4.1743933880915405,
      "val_loss_mixed": 5.300999763684395,
      "validation_loss_regression_vs_previous": false
    },
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_9000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_9000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 9000,
      "cloze_en_contains": 0.02,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.14,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.22789783889980353,
      "distinct_2": 0.43820224719101125,
      "generation_pass_rate": null,
      "generation_pass_rate_regression_vs_previous": false,
      "generation_passed_prompts": 0,
      "generation_scored_prompts": 0,
      "generation_total_prompts": 40,
      "language_consistency_en": 0.875,
      "language_consistency_it": 0.75,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.625,
      "model_type": "pretrained",
      "ppl_en": 171.45126502891384,
      "ppl_it": 66.42193921604036,
      "ppl_mixed": 220.5988020791894,
      "repeated_4gram_rate": 0.85,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 5.264359247116816,
      "source_loss_books_it": 4.98773193359375,
      "source_loss_code": 8.045408121744792,
      "source_loss_web_en": 5.933168110094573,
      "source_loss_web_it": 6.139874508506374,
      "source_loss_wiki_en": 4.055488586425781,
      "source_loss_wiki_it": 3.8951644897460938,
      "val_loss_en": 5.144299057294738,
      "val_loss_it": 4.19602741178919,
      "val_loss_mixed": 5.396345676519932,
      "validation_loss_regression_vs_previous": false
    },
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_10000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 10000,
      "cloze_en_contains": 0.02,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.14,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.22445255474452555,
      "distinct_2": 0.44507575757575757,
      "generation_pass_rate": null,
      "generation_pass_rate_regression_vs_previous": false,
      "generation_passed_prompts": 0,
      "generation_scored_prompts": 0,
      "generation_total_prompts": 40,
      "language_consistency_en": 0.95,
      "language_consistency_it": 0.7625,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.45,
      "model_type": "pretrained",
      "ppl_en": 152.58272957499497,
      "ppl_it": 59.93341541397509,
      "ppl_mixed": 193.9744042501709,
      "repeated_4gram_rate": 0.925,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 4.93456304640997,
      "source_loss_books_it": 4.881648472377232,
      "source_loss_code": 7.710747273763021,
      "source_loss_web_en": 5.753730773925781,
      "source_loss_web_it": 6.29163962916324,
      "source_loss_wiki_en": 4.051776885986328,
      "source_loss_wiki_it": 3.813598937988281,
      "val_loss_en": 5.027706937969856,
      "val_loss_it": 4.093234202900871,
      "val_loss_mixed": 5.2677262135041065,
      "validation_loss_regression_vs_previous": false
    },
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260603_fresh-gpt2small-lr2e4-bs6-wsd-shortfastdecay11k-final5e6-webwiki/step_11000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 11000,
      "cloze_en_contains": 0.02,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.1,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.24277456647398843,
      "distinct_2": 0.49899799599198397,
      "generation_pass_rate": null,
      "generation_pass_rate_regression_vs_previous": false,
      "generation_passed_prompts": 0,
      "generation_scored_prompts": 0,
      "generation_total_prompts": 40,
      "language_consistency_en": 0.95,
      "language_consistency_it": 0.8107142857142857,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.4,
      "model_type": "pretrained",
      "ppl_en": 147.3916153866132,
      "ppl_it": 56.145236622362084,
      "ppl_mixed": 186.23824092873335,
      "repeated_4gram_rate": 0.9,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 4.916778201148624,
      "source_loss_books_it": 4.800248645600819,
      "source_loss_code": 7.810797119140625,
      "source_loss_web_en": 5.808176141036184,
      "source_loss_web_it": 6.264054950914885,
      "source_loss_wiki_en": 4.047422235662287,
      "source_loss_wiki_it": 3.762100830078125,
      "val_loss_en": 4.993093094735776,
      "val_loss_it": 4.0279418445024335,
      "val_loss_mixed": 5.227026719313401,
      "validation_loss_regression_vs_previous": false
    }
  ]
}