{
  "timestamp": "2026-04-29T05:07:06.271644",
  "adapter_path": "./output_ky_bf16_r16_lr2e4_3ep/final_adapter",
  "perplexity": {
    "ky": {
      "loss": 1.6071,
      "ppl": 4.99,
      "n_samples": 500,
      "n_tokens": 128000
    },
    "kz": {
      "loss": 3.7835,
      "ppl": 43.97,
      "n_samples": 500,
      "n_tokens": 113336
    },
    "uz": {
      "loss": 4.0779,
      "ppl": 59.02,
      "n_samples": 500,
      "n_tokens": 127308
    }
  },
  "ner_wikiann": {
    "ky": {
      "precision": 0.109,
      "recall": 0.333,
      "f1": 0.164,
      "n_evaluated": 100,
      "parse_failures": 0
    },
    "kz": {
      "precision": 0.132,
      "recall": 0.375,
      "f1": 0.195,
      "n_evaluated": 100,
      "parse_failures": 0
    },
    "uz": {
      "precision": 0.22,
      "recall": 0.563,
      "f1": 0.316,
      "n_evaluated": 100,
      "parse_failures": 0
    }
  },
  "ner_loglik": {
    "ky": {
      "method": "loglik_span_typing",
      "type_accuracy": 0.586,
      "macro_f1": 0.485,
      "n_spans": 111
    },
    "kz": {
      "method": "loglik_span_typing",
      "type_accuracy": 0.705,
      "macro_f1": 0.543,
      "n_spans": 112
    },
    "uz": {
      "method": "loglik_span_typing",
      "type_accuracy": 0.641,
      "macro_f1": 0.506,
      "n_spans": 103
    }
  },
  "tumlu_qa": {
    "kazakh": {
      "accuracy": 0.3073,
      "correct": 244,
      "n_questions": 794
    },
    "uzbek": {
      "accuracy": 0.2892,
      "correct": 201,
      "n_questions": 695
    },
    "kyrgyz": {
      "accuracy": 0.3608,
      "correct": 267,
      "n_questions": 740
    }
  },
  "_notes": "Merged report. PPL + NER (gen) + NER (loglik) recovered from eval.log.partial1 (first run, killed during TUMLU). TUMLU rerun on 2026-04-29 with --skip_ppl --skip_ner --skip_ner_loglik. Per-type NER breakdown not recoverable from log (only aggregate metrics were printed)."
}