{
  "contains_mixed_model_types": false,
  "metric_recommendations": {
    "cloze_en_contains": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "max",
      "value": 0.08
    },
    "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_14000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260522_fresh-gpt2small-lr1e4-bs6-cosine-webwiki/step_14000.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_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "max",
      "value": 0.2483221476510067
    },
    "distinct_2": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "max",
      "value": 0.4795613160518445
    },
    "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-earlydecay3500-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 0.55
    },
    "ppl_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605211121_fresh-gpt2small-lr2e4-bs6-cosine-webwiki/step_7000.pt",
      "direction": "min",
      "value": 149.78385323563708
    },
    "ppl_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "min",
      "value": 50.61348787793464
    },
    "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_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "min",
      "value": 0.8
    },
    "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/202605211121_fresh-gpt2small-lr2e4-bs6-cosine-webwiki/step_7000.pt",
      "direction": "min",
      "value": 4.845814296177456
    },
    "source_loss_code": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "min",
      "value": 8.125110880533855
    },
    "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_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay7000-final5e6-webwiki/step_7000.pt",
      "direction": "min",
      "value": 4.018826571377841
    },
    "source_loss_wiki_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "min",
      "value": 3.6993704223632813
    },
    "val_loss_en": {
      "checkpoint_name": "step_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605211121_fresh-gpt2small-lr2e4-bs6-cosine-webwiki/step_7000.pt",
      "direction": "min",
      "value": 5.009193276459316
    },
    "val_loss_it": {
      "checkpoint_name": "step_11000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "direction": "min",
      "value": 3.924218099625384
    },
    "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_7000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260530_fresh-gpt2small-lr2e4-bs6-wsd-earlydecay3500-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": 0.2016498625114574,
      "distinct_2": 0.4034253092293054,
      "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": 0.825,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.55,
      "model_type": "pretrained",
      "ppl_en": 173.942553211812,
      "ppl_it": 57.43075774842341,
      "ppl_mixed": 187.87849238225104,
      "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": 4.795341491699219,
      "source_loss_books_it": 4.879483177548363,
      "source_loss_code": 8.230809020996094,
      "source_loss_web_en": 5.886721560829564,
      "source_loss_web_it": 6.098954451711554,
      "source_loss_wiki_en": 4.070300709117543,
      "source_loss_wiki_it": 3.948580627441406,
      "val_loss_en": 5.15872509074661,
      "val_loss_it": 4.050580009085233,
      "val_loss_mixed": 5.235795436761318,
      "validation_loss_regression_vs_previous": false
    },
    {
      "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/202605211121_fresh-gpt2small-lr2e4-bs6-cosine-webwiki/step_7000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 7000,
      "cloze_en_contains": 0.02,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.06,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.210576015108593,
      "distinct_2": 0.43137254901960786,
      "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": 0.775,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.575,
      "model_type": "pretrained",
      "ppl_en": 149.78385323563708,
      "ppl_it": 55.04239975005513,
      "ppl_mixed": 211.84098851476426,
      "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": 4.798703148251488,
      "source_loss_books_it": 4.845814296177456,
      "source_loss_code": 8.518333435058594,
      "source_loss_web_en": 5.82051568282278,
      "source_loss_web_it": 6.103783456902755,
      "source_loss_wiki_en": 4.053226470947266,
      "source_loss_wiki_it": 3.987869873046875,
      "val_loss_en": 5.009193276459316,
      "val_loss_it": 4.008103792784644,
      "val_loss_mixed": 5.355835939065004,
      "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/202605181849_stable-config-recipe-v5-gpt2small-lr2e4-batchmaxpossible-bs6-wsd-fastdecay-webwiki/step_11000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 11000,
      "cloze_en_contains": 0.08,
      "cloze_en_exact": 0.0,
      "cloze_it_contains": 0.12,
      "cloze_it_exact": 0.0,
      "delta_vs_previous_generation_pass_rate": null,
      "delta_vs_previous_validation_loss_mean": null,
      "distinct_1": 0.2483221476510067,
      "distinct_2": 0.4795613160518445,
      "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": 0.8,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.55,
      "model_type": "pretrained",
      "ppl_en": 182.90785540219542,
      "ppl_it": 50.61348787793464,
      "ppl_mixed": 212.20449333464575,
      "repeated_4gram_rate": 0.8,
      "run_dir": null,
      "run_name": null,
      "selected_by": "manual_path",
      "selection_metric_name": null,
      "selection_metric_value": null,
      "source_loss_books_en": 4.827370053245907,
      "source_loss_books_it": 5.0528985886346724,
      "source_loss_code": 8.125110880533855,
      "source_loss_web_en": 6.170190911543997,
      "source_loss_web_it": 6.1338043212890625,
      "source_loss_wiki_en": 4.384110537442294,
      "source_loss_wiki_it": 3.6993704223632813,
      "val_loss_en": 5.208982503639077,
      "val_loss_it": 3.924218099625384,
      "val_loss_mixed": 5.357550400954026,
      "validation_loss_regression_vs_previous": false
    },
    {
      "aggregate_dataset_count": 0,
      "aggregate_validation_loss_mean": null,
      "aggregate_validation_perplexity_mean": null,
      "checkpoint_name": "step_14000",
      "checkpoint_path": "/mnt/apps/llm-nanochat/checkpoints/20260522_fresh-gpt2small-lr1e4-bs6-cosine-webwiki/step_14000.pt",
      "checkpoint_selector": "manual",
      "checkpoint_step": 14000,
      "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": 0.18288288288288287,
      "distinct_2": 0.37383177570093457,
      "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": 0.75,
      "language_switch_rate_en": 0.0,
      "language_switch_rate_it": 0.0,
      "loop_rate": 0.65,
      "model_type": "pretrained",
      "ppl_en": 191.23144990433747,
      "ppl_it": 77.36074444134417,
      "ppl_mixed": 232.58735985024205,
      "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": 5.349181402297247,
      "source_loss_books_it": 5.0058779035295755,
      "source_loss_code": 8.257271321614583,
      "source_loss_web_en": 6.1860496119449015,
      "source_loss_web_it": 6.200174432051809,
      "source_loss_wiki_en": 4.372686212713068,
      "source_loss_wiki_it": 4.148765869140625,
      "val_loss_en": 5.253484474038178,
      "val_loss_it": 4.348479474177126,
      "val_loss_mixed": 5.449265895745693,
      "validation_loss_regression_vs_previous": false
    }
  ]
}