{
  "contains_mixed_model_types": false,
  "metric_recommendations": {
    "cloze_en_contains": {
      "checkpoint_name": "step_8000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_8000.pt",
      "direction": "max",
      "value": 0.06
    },
    "cloze_en_exact": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.0
    },
    "cloze_it_contains": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_10000.pt",
      "direction": "max",
      "value": 0.14
    },
    "cloze_it_exact": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.0
    },
    "distinct_1": {
      "checkpoint_name": "step_9000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_9000.pt",
      "direction": "max",
      "value": 0.21669793621013134
    },
    "distinct_2": {
      "checkpoint_name": "step_9000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_9000.pt",
      "direction": "max",
      "value": 0.4395711500974659
    },
    "language_consistency_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 1.0
    },
    "language_consistency_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "max",
      "value": 0.85
    },
    "language_switch_rate_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-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/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 0.0
    },
    "loop_rate": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 0.575
    },
    "ppl_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 151.13060966318096
    },
    "ppl_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 53.415595943431654
    },
    "ppl_mixed": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 184.16156916662413
    },
    "repeated_4gram_rate": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_10000.pt",
      "direction": "min",
      "value": 0.85
    },
    "source_loss_books_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 4.756146385556176
    },
    "source_loss_books_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 4.887685866582961
    },
    "source_loss_code": {
      "checkpoint_name": "step_9000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_9000.pt",
      "direction": "min",
      "value": 7.898111470540365
    },
    "source_loss_web_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 5.799130891498766
    },
    "source_loss_web_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 6.034188923082854
    },
    "source_loss_wiki_en": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_10000.pt",
      "direction": "min",
      "value": 3.9518307772549717
    },
    "source_loss_wiki_it": {
      "checkpoint_name": "step_10000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_10000.pt",
      "direction": "min",
      "value": 3.7821286010742186
    },
    "val_loss_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 5.018144427605395
    },
    "val_loss_it": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 3.978102762191022
    },
    "val_loss_mixed": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 5.215813465607472
    }
  },
  "recommended_checkpoint": {
    "checkpoint_name": "step_7000",
    "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
    "direction": "min",
    "value": 5.215813465607472
  },
  "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/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 7000,
      "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": 0.19534883720930232,
      "distinct_2": 0.4067632850241546,
      "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": 1.0,
      "language_consistency_it": 0.85,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.575,
      "model_type": "pretrained",
      "ppl_en": 151.13060966318096,
      "ppl_it": 53.415595943431654,
      "ppl_mixed": 184.16156916662413,
      "repeated_4gram_rate": 0.95,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 4.756146385556176,
      "source_loss_books_it": 4.887685866582961,
      "source_loss_code": 8.180503336588542,
      "source_loss_web_en": 5.799130891498766,
      "source_loss_web_it": 6.034188923082854,
      "source_loss_wiki_en": 4.018826571377841,
      "source_loss_wiki_it": 4.058106689453125,
      "val_loss_en": 5.018144427605395,
      "val_loss_it": 3.978102762191022,
      "val_loss_mixed": 5.215813465607472,
      "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/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_8000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 8000,
      "cloze_en_contains": 0.06,
      "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": 0.19275929549902152,
      "distinct_2": 0.38250254323499494,
      "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": 0.975,
      "language_consistency_it": 0.85,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.65,
      "model_type": "pretrained",
      "ppl_en": 212.9131475472987,
      "ppl_it": 69.86212260454919,
      "ppl_mixed": 201.18270985357594,
      "repeated_4gram_rate": 0.95,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 5.166426885695684,
      "source_loss_books_it": 5.114150637672061,
      "source_loss_code": 8.134779866536459,
      "source_loss_web_en": 5.994843733938117,
      "source_loss_web_it": 6.211712485865543,
      "source_loss_wiki_en": 4.121847326105291,
      "source_loss_wiki_it": 4.107617797851563,
      "val_loss_en": 5.360884324559626,
      "val_loss_it": 4.246523622606621,
      "val_loss_mixed": 5.304213499411558,
      "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/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_9000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 9000,
      "cloze_en_contains": 0.02,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.08,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.21669793621013134,
      "distinct_2": 0.4395711500974659,
      "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": 0.925,
      "language_consistency_it": 0.825,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.575,
      "model_type": "pretrained",
      "ppl_en": 203.44163559118488,
      "ppl_it": 68.00750896530816,
      "ppl_mixed": 254.88522562127972,
      "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": 5.315783546084449,
      "source_loss_books_it": 4.982829139346168,
      "source_loss_code": 7.898111470540365,
      "source_loss_web_en": 6.196750440095601,
      "source_loss_web_it": 6.381468120374177,
      "source_loss_wiki_en": 3.9779129028320312,
      "source_loss_wiki_it": 4.027180786132813,
      "val_loss_en": 5.315379160755086,
      "val_loss_it": 4.219618125040023,
      "val_loss_mixed": 5.540813348232171,
      "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/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_10000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 10000,
      "cloze_en_contains": 0.04,
      "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": 0.20783956244302643,
      "distinct_2": 0.4011352885525071,
      "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": 0.975,
      "language_consistency_it": 0.775,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.05,
      "loop_rate": 0.675,
      "model_type": "pretrained",
      "ppl_en": 214.2336935220254,
      "ppl_it": 71.46403426847351,
      "ppl_mixed": 272.798798629929,
      "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": 5.389402843656994,
      "source_loss_books_it": 5.126949128650484,
      "source_loss_code": 8.170843505859375,
      "source_loss_web_en": 6.187845330489309,
      "source_loss_web_it": 6.791255750154194,
      "source_loss_wiki_en": 3.9518307772549717,
      "source_loss_wiki_it": 3.7821286010742186,
      "val_loss_en": 5.367067444999263,
      "val_loss_it": 4.26919430592021,
      "val_loss_mixed": 5.608734522110376,
      "validation_loss_regression_vs_previous": false
    }
  ]
}