{
 "date_long": "25 October 2026",
 "models": {
  "en": {
   "release_id": "flyingkiwi-en-3-1",
   "params_m": 527.63136,
   "vocab": 85121,
   "token_increase": 0.0023277321503793586,
   "download_mb": 571.378688,
   "model_mb": 560.82336,
   "weight_bits": 8,
   "kl": 0.011447516499902121,
   "top1": 0.9666666666666667,
   "kl_wasm_relaxed": 0.014706084268012395,
   "kl_wasm_simd": 0.012604743865085765,
   "history": {
    "lang": "en",
    "train_samples": 15834,
    "val_samples": 344,
    "steps": 148,
    "qat_start": 148,
    "lr": 1.5e-05,
    "micro_tokens": 16384,
    "accum": 4,
    "log": [
     {
      "step": 10,
      "loss": 1.5626052643940465,
      "lr": 1.4848497862896882e-05,
      "gnorm": 4.200143814086914,
      "target_tokens": 313864,
      "minutes": 0.4809096097946167
     },
     {
      "step": 20,
      "loss": 1.4452302825240568,
      "lr": 1.4400792310119917e-05,
      "gnorm": 2.5857977867126465,
      "target_tokens": 577735,
      "minutes": 0.9570551077524821
     },
     {
      "step": 30,
      "loss": 1.3499130592316597,
      "lr": 1.367698062997685e-05,
      "gnorm": 1.8446320295333862,
      "target_tokens": 899116,
      "minutes": 1.437689193089803
     },
     {
      "step": 40,
      "loss": 1.4254186515112794,
      "lr": 1.270955438288305e-05,
      "gnorm": 2.2187139987945557,
      "target_tokens": 1227241,
      "minutes": 1.9237019459406535
     },
     {
      "step": 50,
      "loss": 1.4402376408578699,
      "lr": 1.1541940870742035e-05,
      "gnorm": 2.1185367107391357,
      "target_tokens": 1513096,
      "minutes": 2.3925805568695067
     },
     {
      "step": 60,
      "loss": 1.4259786586523415,
      "lr": 1.0226553706116715e-05,
      "gnorm": 2.055525779724121,
      "target_tokens": 1809915,
      "minutes": 2.875098959604899
     },
     {
      "step": 70,
      "loss": 1.4854095034035146,
      "lr": 8.822439990209688e-06,
      "gnorm": 1.8960177898406982,
      "target_tokens": 2134384,
      "minutes": 3.3578696727752684
     },
     {
      "step": 80,
      "loss": 1.2391044620129923,
      "lr": 7.3926297167085684e-06,
      "gnorm": 1.921399712562561,
      "target_tokens": 2439846,
      "minutes": 3.826851749420166
     },
     {
      "step": 90,
      "loss": 1.1370692397935112,
      "lr": 6.001306385491112e-06,
      "gnorm": 1.990168809890747,
      "target_tokens": 2760725,
      "minutes": 4.312385110060374
     },
     {
      "step": 100,
      "loss": 1.166401613017478,
      "lr": 4.710925835988687e-06,
      "gnorm": 1.8513766527175903,
      "target_tokens": 3053583,
      "minutes": 4.779878393809001,
      "val": 1.3245368406847624
     },
     {
      "step": 110,
      "loss": 1.1604635997828667,
      "lr": 3.579412634401564e-06,
      "gnorm": 2.0154128074645996,
      "target_tokens": 3335248,
      "minutes": 5.270262801647187
     },
     {
      "step": 120,
      "loss": 1.2641661743524522,
      "lr": 2.657559867604066e-06,
      "gnorm": 2.0432775020599365,
      "target_tokens": 3660880,
      "minutes": 5.7549683332443236
     },
     {
      "step": 130,
      "loss": 1.1707798459964058,
      "lr": 1.986749065748562e-06,
      "gnorm": 1.9341208934783936,
      "target_tokens": 3991108,
      "minutes": 6.244042952855428
     },
     {
      "step": 140,
      "loss": 1.1327010649179097,
      "lr": 1.597092605152968e-06,
      "gnorm": 2.6037070751190186,
      "target_tokens": 4301989,
      "minutes": 6.731195823351542
     },
     {
      "step": 148,
      "loss": 1.2334547916053618,
      "lr": 1.5e-06,
      "gnorm": 2.7293732166290283,
      "target_tokens": 4513064,
      "minutes": 7.094963733355204,
      "val": 1.3112765079725426
     }
    ],
    "val_initial": 2.164960640848029,
    "val_final": 1.3112765079725426,
    "qat": false,
    "minutes": 7.146401607990265
   },
   "rewrite": {
    "answers": {
     "rewritten": 22067,
     "longer": 24,
     "lost_code": 759,
     "self_disclaimer": 492,
     "language_or_format": 5,
     "missing": 2,
     "thinking_aloud": 3
    },
    "characters_before": 22614178,
    "characters_after": 9683327,
    "kept_share": 0.428
   }
  },
  "tr": {
   "release_id": "flyingkiwi-tr-3-1",
   "params_m": 520.600576,
   "vocab": 78255,
   "token_increase": 0.0021155076811535233,
   "download_mb": 566.790616,
   "model_mb": 553.353152,
   "weight_bits": 8,
   "kl": 0.051268519298173484,
   "top1": 0.9666666666666667,
   "kl_wasm_relaxed": 0.05633755976644655,
   "kl_wasm_simd": 0.05612133260195454,
   "history": {
    "lang": "tr",
    "train_samples": 20770,
    "val_samples": 388,
    "steps": 300,
    "qat_start": 300,
    "lr": 1.5e-05,
    "micro_tokens": 16384,
    "accum": 4,
    "log": [
     {
      "step": 10,
      "loss": 2.2331537584699794,
      "lr": 1.0972883382072952e-05,
      "gnorm": 6.048584461212158,
      "target_tokens": 321192,
      "minutes": 0.4915636936823527
     },
     {
      "step": 20,
      "loss": 1.971697739001885,
      "lr": 1.4852496304953189e-05,
      "gnorm": 2.9622976779937744,
      "target_tokens": 658761,
      "minutes": 0.9835285226504008
     },
     {
      "step": 30,
      "loss": 2.030714451579121,
      "lr": 1.4669631484992288e-05,
      "gnorm": 2.4321014881134033,
      "target_tokens": 993552,
      "minutes": 1.4695297559102376
     },
     {
      "step": 40,
      "loss": 1.9536323610734923,
      "lr": 1.4416431839087556e-05,
      "gnorm": 2.8138365745544434,
      "target_tokens": 1333880,
      "minutes": 1.976205587387085
     },
     {
      "step": 50,
      "loss": 1.827752637387279,
      "lr": 1.4095671475544961e-05,
      "gnorm": 2.108384370803833,
      "target_tokens": 1670239,
      "minutes": 2.4627658565839132
     },
     {
      "step": 60,
      "loss": 1.8947722679938035,
      "lr": 1.3710864712030896e-05,
      "gnorm": 2.2460174560546875,
      "target_tokens": 2022028,
      "minutes": 2.968234062194824
     },
     {
      "step": 70,
      "loss": 1.8994885522317282,
      "lr": 1.3266227571972412e-05,
      "gnorm": 2.499967098236084,
      "target_tokens": 2358417,
      "minutes": 3.4648415009180704
     },
     {
      "step": 80,
      "loss": 1.7908881940539618,
      "lr": 1.2766631592922294e-05,
      "gnorm": 2.499741554260254,
      "target_tokens": 2720906,
      "minutes": 3.9803888320922853
     },
     {
      "step": 90,
      "loss": 1.801835294764868,
      "lr": 1.2217550452974194e-05,
      "gnorm": 2.347470998764038,
      "target_tokens": 3023951,
      "minutes": 4.458623381455739
     },
     {
      "step": 100,
      "loss": 1.6783337099151943,
      "lr": 1.1625000000000001e-05,
      "gnorm": 2.4776923656463623,
      "target_tokens": 3356097,
      "minutes": 4.958300260702769,
      "val": 1.7490525836296371
     },
     {
      "step": 110,
      "loss": 1.722389840031328,
      "lr": 1.0995472340761651e-05,
      "gnorm": 2.2078847885131836,
      "target_tokens": 3698135,
      "minutes": 5.511547911167145
     },
     {
      "step": 120,
      "loss": 1.7449805838059373,
      "lr": 1.0335864712030895e-05,
      "gnorm": 2.093627691268921,
      "target_tokens": 4040920,
      "minutes": 5.999652886390686
     },
     {
      "step": 130,
      "loss": 1.5988771505621788,
      "lr": 9.653403913019875e-06,
      "gnorm": 1.969544768333435,
      "target_tokens": 4393068,
      "minutes": 6.511383970578511
     },
     {
      "step": 140,
      "loss": 1.8443494129260183,
      "lr": 8.95556712705666e-06,
      "gnorm": 2.5523858070373535,
      "target_tokens": 4728987,
      "minutes": 7.000515540440877
     },
     {
      "step": 150,
      "loss": 1.74615673465753,
      "lr": 8.25e-06,
      "gnorm": 2.4716780185699463,
      "target_tokens": 5069457,
      "minutes": 7.493805392583211
     },
     {
      "step": 160,
      "loss": 1.5048379656279582,
      "lr": 7.5444328729433405e-06,
      "gnorm": 1.9842160940170288,
      "target_tokens": 5387520,
      "minutes": 7.990439907709757
     },
     {
      "step": 170,
      "loss": 1.5442331042146855,
      "lr": 6.846596086980127e-06,
      "gnorm": 2.0844287872314453,
      "target_tokens": 5702812,
      "minutes": 8.472988402843475
     },
     {
      "step": 180,
      "loss": 1.5646914317703828,
      "lr": 6.164135287969106e-06,
      "gnorm": 2.1184427738189697,
      "target_tokens": 6060644,
      "minutes": 8.981649311383565
     },
     {
      "step": 190,
      "loss": 1.4684271788551988,
      "lr": 5.50452765923835e-06,
      "gnorm": 1.9899039268493652,
      "target_tokens": 6401735,
      "minutes": 9.48187431494395
     },
     {
      "step": 200,
      "loss": 1.506901257947503,
      "lr": 4.875000000000002e-06,
      "gnorm": 1.9435840845108032,
      "target_tokens": 6736678,
      "minutes": 9.98393474817276,
      "val": 1.659642864784327
     },
     {
      "step": 210,
      "loss": 1.6127409248942086,
      "lr": 4.2824495470258075e-06,
      "gnorm": 2.115579843521118,
      "target_tokens": 7073529,
      "minutes": 10.526639445622761
     },
     {
      "step": 220,
      "loss": 1.5735964759503667,
      "lr": 3.7333684070777093e-06,
      "gnorm": 2.314983606338501,
      "target_tokens": 7405816,
      "minutes": 11.013392003377279
     },
     {
      "step": 230,
      "loss": 1.5383553986537837,
      "lr": 3.2337724280275885e-06,
      "gnorm": 2.054607629776001,
      "target_tokens": 7766658,
      "minutes": 11.522872463862102
     },
     {
      "step": 240,
      "loss": 1.5980819798265093,
      "lr": 2.789135287969106e-06,
      "gnorm": 2.249545097351074,
      "target_tokens": 8094124,
      "minutes": 12.007297496000925
     },
     {
      "step": 250,
      "loss": 1.4667420361461099,
      "lr": 2.404328524455039e-06,
      "gnorm": 2.104658842086792,
      "target_tokens": 8417501,
      "minutes": 12.50293847322464
     },
     {
      "step": 260,
      "loss": 1.5177092329693314,
      "lr": 2.0835681609124438e-06,
      "gnorm": 1.9299768209457397,
      "target_tokens": 8757085,
      "minutes": 12.991623536745708
     },
     {
      "step": 270,
      "loss": 1.4815499805433825,
      "lr": 1.8303685150077138e-06,
      "gnorm": 2.033214807510376,
      "target_tokens": 9083782,
      "minutes": 13.484034279982248
     },
     {
      "step": 280,
      "loss": 1.4673382154133239,
      "lr": 1.6475036950468118e-06,
      "gnorm": 2.109820604324341,
      "target_tokens": 9460517,
      "minutes": 13.989264289538065
     },
     {
      "step": 290,
      "loss": 1.549758317701472,
      "lr": 1.5369772062641554e-06,
      "gnorm": 1.9484636783599854,
      "target_tokens": 9792928,
      "minutes": 14.483213527997334
     },
     {
      "step": 300,
      "loss": 1.5063842165939338,
      "lr": 1.5e-06,
      "gnorm": 2.2882871627807617,
      "target_tokens": 10138914,
      "minutes": 14.984930046399434,
      "val": 1.6324840817759554
     }
    ],
    "val_initial": 2.6102495054579546,
    "val_final": 1.6324840817759554,
    "qat": false,
    "minutes": 15.071084189414979
   },
   "rewrite": {
    "answers": {
     "rewritten": 30036,
     "language_or_format": 87,
     "lost_code": 1176,
     "self_disclaimer": 227,
     "longer": 25,
     "missing": 35
    },
    "characters_before": 45715864,
    "characters_after": 15113306,
    "kept_share": 0.331
   }
  }
 },
 "frozen": {
  "en": {
   "reference": {
    "answerable": {
     "correct": 22,
     "total": 30
    },
    "all": {
     "correct": 63,
     "total": 100
    },
    "unknowable": {
     "correct": 17,
     "total": 20
    },
    "false_premise": {
     "correct": 7,
     "total": 10
    },
    "ambiguity": {
     "correct": 0,
     "total": 10
    },
    "friendly": {
     "correct": 6,
     "total": 10
    },
    "multi_turn": {
     "correct": 11,
     "total": 20
    },
    "flags": {
     "language_errors": 0,
     "fabrications": 6
    }
   },
   "flyingkiwi": {
    "answerable": {
     "correct": 20,
     "total": 30
    },
    "all": {
     "correct": 63,
     "total": 100
    },
    "unknowable": {
     "correct": 18,
     "total": 20
    },
    "false_premise": {
     "correct": 5,
     "total": 10
    },
    "ambiguity": {
     "correct": 3,
     "total": 10
    },
    "friendly": {
     "correct": 6,
     "total": 10
    },
    "multi_turn": {
     "correct": 11,
     "total": 20
    },
    "flags": {
     "language_errors": 0,
     "fabrications": 6
    }
   }
  },
  "tr": {
   "reference": {
    "answerable": {
     "correct": 9,
     "total": 30
    },
    "all": {
     "correct": 12,
     "total": 100
    },
    "unknowable": {
     "correct": 1,
     "total": 20
    },
    "false_premise": {
     "correct": 0,
     "total": 10
    },
    "ambiguity": {
     "correct": 0,
     "total": 10
    },
    "friendly": {
     "correct": 0,
     "total": 10
    },
    "multi_turn": {
     "correct": 2,
     "total": 20
    },
    "flags": {
     "language_errors": 0,
     "fabrications": 19
    }
   },
   "flyingkiwi": {
    "answerable": {
     "correct": 11,
     "total": 30
    },
    "all": {
     "correct": 37,
     "total": 100
    },
    "unknowable": {
     "correct": 14,
     "total": 20
    },
    "false_premise": {
     "correct": 2,
     "total": 10
    },
    "ambiguity": {
     "correct": 3,
     "total": 10
    },
    "friendly": {
     "correct": 1,
     "total": 10
    },
    "multi_turn": {
     "correct": 6,
     "total": 20
    },
    "flags": {
     "language_errors": 0,
     "fabrications": 17
    }
   }
  }
 },
 "lengths": {
  "en": {
   "reference": 37.4,
   "v3": 82.8,
   "v3_1": 51.3
  },
  "tr": {
   "reference": 73.8,
   "v3": 255.7,
   "v3_1": 91.5
  }
 },
 "huggingface": {
  "en": "https://huggingface.co/yunusemrejr/Flyingkiwi-EN-0.6B",
  "tr": "https://huggingface.co/yunusemrejr/Flyingkiwi-TR-0.6B"
 },
 "truthfulqa": {
  "reference": 0.2582619339045288,
  "v3_1": 0.25091799265605874,
  "v3": 0.21297429620563035
 },
 "quant": {
  "plan-en-v3c": {
   "student": "training/output/kiwi-v3c/student_en",
   "lang": "en",
   "held_out": {
    "file": "sft_en_val.jsonl",
    "conversations": 40,
    "max_tokens": 640
   },
   "calibration": {
    "file": "sft_en_train.jsonl",
    "conversations": 16
   },
   "variants": {
    "float": {
     "answer_nll": 1.32686,
     "tokens": 4279,
     "bytes_estimate": null
    },
    "q8": {
     "answer_nll": 1.32714,
     "tokens": 4279,
     "bytes_estimate": 560843192
    },
    "q4_rtn": {
     "answer_nll": 1.4956,
     "tokens": 4279,
     "bytes_estimate": 340642232
    },
    "q4_full": {
     "answer_nll": 1.48408,
     "tokens": 4279,
     "bytes_estimate": 340642232
    },
    "plan_380": {
     "answer_nll": 1.40408,
     "tokens": 4279,
     "bytes_estimate": 379963832
    }
   },
   "chosen_q4": "q4_full",
   "minutes": 29.3,
   "plans": {
    "380": {
     "q8": 52,
     "bytes_estimate": 379963832,
     "q4": "q4_full"
    }
   }
  },
  "plan-en-v3c-gptq": {
   "student": "training/output/kiwi-v3c/student_en",
   "lang": "en",
   "held_out": {
    "file": "sft_en_val.jsonl",
    "conversations": 40,
    "max_tokens": 640
   },
   "calibration": {
    "file": "sft_en_train.jsonl",
    "conversations": 16
   },
   "variants": {
    "q4_gptq": {
     "answer_nll": 1.40719,
     "tokens": 4279,
     "bytes_estimate": 340642232
    },
    "plan_380": {
     "answer_nll": 1.37595,
     "tokens": 4279,
     "bytes_estimate": 379963832
    },
    "plan_420": {
     "answer_nll": 1.36543,
     "tokens": 4279,
     "bytes_estimate": 419809720
    }
   },
   "chosen_q4": "q4_gptq",
   "minutes": 6.8,
   "plans": {
    "380": {
     "q8": 53,
     "bytes_estimate": 379963832,
     "q4": "q4_gptq"
    },
    "420": {
     "q8": 89,
     "bytes_estimate": 419809720,
     "q4": "q4_gptq"
    }
   }
  },
  "plan-tr-v3c-gptq": {
   "student": "training/output/kiwi-v3c/student_tr",
   "lang": "tr",
   "held_out": {
    "file": "sft_tr_val.jsonl",
    "conversations": 24,
    "max_tokens": 640
   },
   "calibration": {
    "file": "sft_tr_train.jsonl",
    "conversations": 12
   },
   "variants": {
    "float": {
     "answer_nll": 1.56241,
     "tokens": 4598,
     "bytes_estimate": null
    },
    "q8": {
     "answer_nll": 1.5649,
     "tokens": 4598,
     "bytes_estimate": 553372984
    },
    "q4_rtn": {
     "answer_nll": 1.80424,
     "tokens": 4598,
     "bytes_estimate": 333172024
    },
    "q4_gptq": {
     "answer_nll": 1.66387,
     "tokens": 4598,
     "bytes_estimate": 333172024
    },
    "plan_420": {
     "answer_nll": 1.59986,
     "tokens": 4598,
     "bytes_estimate": 419679544
    }
   },
   "chosen_q4": "q4_gptq",
   "minutes": 8.3,
   "plans": {
    "420": {
     "q8": 91,
     "bytes_estimate": 419679544,
     "q4": "q4_gptq"
    }
   }
  },
  "tr-v3b-formats": {
   "model": "Turkish v3b fine-tune (training/output/kiwi-v3/release_tr), trained with 4-bit quantisation-aware training in its last 30% (steps 439-628)",
   "held_out": "48 conversations from sft_tr_val.jsonl (validation split), at most 640 tokens each, assistant tokens only",
   "metric": "mean negative log-likelihood of the assistant tokens (nats/token), PyTorch with dequantised weights",
   "variants": {
    "float": 1.5393,
    "q8 (max|w|/127 per 32)": 1.5401,
    "q4 plain rounding (max|w|/7 per 32)": 1.5425,
    "q4 activation-weighted clipped scale (0.5-1.0 x max|w|/7)": 1.6259,
    "q4 GPTQ (128 calibration conversations, damp 0.01, block 128)": 1.5965,
    "plan 340 MB: q4 clipped + 13 Fisher-ranked matrices at q8": 1.6124
   },
   "bytes": {
    "q8": 553372984,
    "q4": 333172024
   },
   "earlier_check": "12 held-out conversations: float 1.5259, plain q4 1.5375, GPTQ 1.5688; GPTQ restricted to attention, MLP, layers 0-13, layers 14-27, q/k or v/o all scored worse than plain rounding (1.5411-1.5656).",
   "reading": "This model was trained with 4-bit QAT using exactly the plain max-scaled rounding, so its weights sit on that grid: plain rounding costs 0.0024 nats over 8 bits, while GPTQ (60-80% lower output error per layer on held-out activations) and clipped scales move weights off the trained grid and cost 0.06-0.09 nats. The result says nothing about models trained without QAT: for the English v3c fine-tune (no QAT) plain 4-bit rounding costs 0.17 nats (evidence/eval-v3/quant/plan-en-v3c.json)."
  }
 },
 "small_device": {
  "method": "tools/e2e_browser.mjs (fresh headless Chrome 155 profile, real UI) inside a systemd user scope (CPUQuota, MemoryMax, taskset); bundles served from 127.0.0.1 by tools/serve_dev.py with production headers. Emulation, not a real phone: CPU quotas slow every core evenly and do not reproduce ARM cores, thermal limits or mobile browsers.",
  "profiles": {
   "laptop": "AMD Ryzen 5 7530U laptop, no limits; the page sees 12 cores and the app starts 6 engine threads",
   "uncapped-2threads": "the page sees 4 cores (app starts 2 engine threads), 4 physical cores and 200% CPU allowed (does not slow 2 threads)",
   "mid": "mid-range phone emulation: the page sees 4 cores (2 engine threads); Chrome limited to 4 physical cores and 100% CPU in total (each engine thread at about half a laptop core) and 2 GB of memory; 390x844 phone viewport",
   "low": "low-end phone emulation: the page sees 2 cores (1 engine thread); 2 physical cores, 25% CPU in total (about a quarter of a laptop core), 1.5 GB of memory; phone viewport"
  },
  "messages": [
   "palindrome function in Python",
   "follow-up: ignore spaces and capital letters"
  ],
  "runs": [
   {
    "release": "flyingkiwi-en-3-0",
    "profile": "low",
    "engine": "C → WebAssembly relaxed SIMD · 1 thread · 8-bit weights",
    "first_reply_seconds_including_load": 79.196,
    "peak_memory_mb": 1122,
    "replies": [
     {
      "seconds": 76.004,
      "meta": "134 tokens · 3.3 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 86.69,
      "meta": "225 tokens · 2.7 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-en-3-0",
    "profile": "mid",
    "engine": "C → WebAssembly relaxed SIMD · 2 threads · 8-bit weights",
    "first_reply_seconds_including_load": 26.106,
    "peak_memory_mb": 1288,
    "replies": [
     {
      "seconds": 24.214,
      "meta": "144 tokens · 8.6 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 23.183,
      "meta": "182 tokens · 8.2 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-en-3-0",
    "profile": "uncapped-2threads",
    "engine": "C → WebAssembly relaxed SIMD · 2 threads · 8-bit weights",
    "first_reply_seconds_including_load": 14.862,
    "peak_memory_mb": 1266,
    "replies": [
     {
      "seconds": 13.056,
      "meta": "118 tokens · 17.6 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 16.075,
      "meta": "210 tokens · 14.2 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-en-3-1",
    "profile": "laptop",
    "engine": "C → WebAssembly relaxed SIMD · 6 threads · 8-bit weights",
    "first_reply_seconds_including_load": 11.838,
    "peak_memory_mb": 1199,
    "replies": [
     {
      "seconds": 10.039,
      "meta": "84 tokens · 21.9 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 5.025,
      "meta": "74 tokens · 18.1 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-en-3-1",
    "profile": "low",
    "engine": "C → WebAssembly relaxed SIMD · 1 thread · 8-bit weights",
    "first_reply_seconds_including_load": 83.204,
    "peak_memory_mb": 1118,
    "replies": [
     {
      "seconds": 79.712,
      "meta": "82 tokens · 2.4 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 64.479,
      "meta": "125 tokens · 2.0 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-en-3-1",
    "profile": "mid",
    "engine": "C → WebAssembly relaxed SIMD · 2 threads · 8-bit weights",
    "first_reply_seconds_including_load": 22.199,
    "peak_memory_mb": 1358,
    "replies": [
     {
      "seconds": 20.286,
      "meta": "69 tokens · 6.0 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 12.101,
      "meta": "74 tokens · 6.5 tokens/s on this device",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-tr-3-1",
    "profile": "laptop",
    "engine": "C → WebAssembly relaxed SIMD · 6 threads · 8-bit weights",
    "first_reply_seconds_including_load": 16.344,
    "peak_memory_mb": 1196,
    "replies": [
     {
      "seconds": 13.068,
      "meta": "140 token · bu cihazda saniyede 20.0 token",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 11.056,
      "meta": "187 token · bu cihazda saniyede 18.5 token",
      "code_blocks": 2,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   },
   {
    "release": "flyingkiwi-tr-3-1",
    "profile": "low",
    "engine": "C → WebAssembly relaxed SIMD · 1 thread · 8-bit weights",
    "first_reply_seconds_including_load": 79.8,
    "peak_memory_mb": 1113,
    "replies": [
     {
      "seconds": 74.902,
      "meta": "59 token · bu cihazda saniyede 2.1 token",
      "code_blocks": 1,
      "error": null
     },
     {
      "seconds": 52.09,
      "meta": "93 token · bu cihazda saniyede 2.0 token",
      "code_blocks": 1,
      "error": null
     }
    ],
    "isolated": true,
    "exit_code": 0
   }
  ]
 },
 "data": {
  "languages": {
   "en": {
    "train": 15834,
    "val": 344,
    "groups": {
     "topics": 6292,
     "human": 3369,
     "honesty": 1890,
     "calibration_unknown": 1426,
     "chained": 1198,
     "gsm8k": 653,
     "calibration_known": 498,
     "fictional": 479,
     "memory": 248,
     "identity": 125
    },
    "followups": 1544,
    "no_system": 2385,
    "target_tokens": 10314874,
    "steps": 240,
    "steps_done": 240,
    "qat": true,
    "qat_start": 168,
    "minutes": 12.638742359479268,
    "val_initial": 2.1528877050135975,
    "val_final": 1.3952952882340877,
    "lr": 1.5e-05
   },
   "tr": {
    "train": 20778,
    "val": 388,
    "groups": {
     "human": 7761,
     "topics": 7455,
     "honesty": 1941,
     "chained": 1567,
     "calibration_unknown": 697,
     "memory": 474,
     "fictional": 460,
     "gsm8k": 433,
     "calibration_known": 260,
     "identity": 118
    },
    "followups": 2393,
    "no_system": 3230,
    "target_tokens": 30271080,
    "steps": 628,
    "steps_done": 628,
    "qat": true,
    "qat_start": 439,
    "minutes": 38.63355012734731,
    "val_initial": 3.0833059536514926,
    "val_final": 1.54861929904373,
    "lr": 1.5e-05
   }
  },
  "summary": "English: 15,834 training and 344 validation conversations: 3,369 human-written prompts (OASST2, Aya), 653 GSM8K word problems, 6,292 teacher-written everyday, explanation, writing, coding and maths requests, 1,890 honesty probes, 479 questions about invented entities, 125 identity questions, 248 memory checks and 1,198 chained conversations; 1,544 continue with teacher-written follow-up turns and 2,385 have no system prompt. Turkish: 20,778 training and 388 validation conversations: 7,761 human-written prompts (OASST2, Aya, translated OASST2), 433 GSM8K word problems, 7,455 teacher-written everyday, explanation, writing, coding and maths requests, 1,941 honesty probes, 460 questions about invented entities, 118 identity questions, 474 memory checks and 1,567 chained conversations; 2,393 continue with teacher-written follow-up turns and 3,230 have no system prompt.",
  "calibration": "TriviaQA questions were put to the unchanged Qwen3-0.6B (greedy plus four samples) in each language. English: 498 questions the base model reliably answered keep a short correct answer, 1,426 it reliably missed get an honest “not sure” reply; Turkish: 260 questions the base model reliably answered keep a short correct answer, 697 it reliably missed get an honest “not sure” reply. Questions in between were left out.",
  "recipe": "Full fine-tuning in BF16 autocast with FP32 master weights, AdamW (β 0.9/0.95), peak learning rate 1.5e-05 with warm-up and cosine decay to 10%, gradient clipping 1.0, loss on assistant tokens only, 4-bit quantisation-aware training for the final 30% of steps. English 240 steps, 10.3M assistant tokens, 13 min, validation loss 2.153 → 1.395; Turkish 628 steps, 30.3M assistant tokens, 39 min, validation loss 3.083 → 1.549.",
  "compute": "One NVIDIA RTX PRO 6000 Blackwell (96 GB) on the existing Colab Pro+ plan: 5.3 h in total, of which 2.6 h of teacher generation; the rest probing, vocabulary building, training and packaging. Evaluation ran locally on a CPU."
 },
 "variants": [
  {
   "lang": "en",
   "variant": "v3a fine-tune, 4-bit (QAT)",
   "shipped": false,
   "reference_all": 64,
   "variant_all": 58,
   "delta": -6,
   "reference_fabrications": 7,
   "variant_fabrications": 11,
   "scores": "v3a/en-scores.json"
  },
  {
   "lang": "en",
   "variant": "v3b fine-tune (cleaned data), 4-bit (QAT)",
   "shipped": false,
   "reference_all": 66,
   "variant_all": 57,
   "delta": -9,
   "reference_fabrications": 8,
   "variant_fabrications": 18,
   "scores": "diag/scores-en-v3b.json"
  },
  {
   "lang": "en",
   "variant": "Qwen3-0.6B without fine-tuning, 4-bit",
   "shipped": false,
   "reference_all": 64,
   "variant_all": 60,
   "delta": -4,
   "reference_fabrications": 6,
   "variant_fabrications": 10,
   "scores": "diag/scores-diag-base-q4.json"
  },
  {
   "lang": "en",
   "variant": "v3b, 8-bit attention + 4-bit feed-forward",
   "shipped": false,
   "reference_all": 65,
   "variant_all": 58,
   "delta": -7,
   "reference_fabrications": 8,
   "variant_fabrications": 15,
   "scores": "diag/scores-diag-v3b-mixed.json"
  },
  {
   "lang": "en",
   "variant": "v3b, 8-bit",
   "shipped": false,
   "reference_all": 65,
   "variant_all": 66,
   "delta": 1,
   "reference_fabrications": 6,
   "variant_fabrications": 13,
   "scores": "diag/scores-diag-v3b-q8.json"
  },
  {
   "lang": "en",
   "variant": "v3b blended 50/50 with Qwen3-0.6B, 8-bit",
   "shipped": true,
   "reference_all": 62,
   "variant_all": 65,
   "delta": 3,
   "reference_fabrications": 6,
   "variant_fabrications": 11,
   "scores": "en-scores.json"
  },
  {
   "lang": "tr",
   "variant": "v3b fine-tune (cleaned data), 8-bit",
   "shipped": true,
   "reference_all": 12,
   "variant_all": 32,
   "delta": 20,
   "reference_fabrications": 26,
   "variant_fabrications": 28,
   "scores": "tr-scores.json"
  },
  {
   "lang": "tr",
   "variant": "v3b blended 50/50 with Qwen3-0.6B, 8-bit",
   "shipped": false,
   "reference_all": 11,
   "variant_all": 18,
   "delta": 7,
   "reference_fabrications": 28,
   "variant_fabrications": 26,
   "scores": "diag/scores-tr-wise50.json"
  },
  {
   "lang": "en",
   "variant": "v3b blended 25/75 with Qwen3-0.6B (25% fine-tune), 8-bit",
   "shipped": false,
   "reference_all": 64,
   "variant_all": 53,
   "delta": -11,
   "reference_fabrications": 7,
   "variant_fabrications": 12,
   "scores": "diag/scores-diag-wise25.json"
  },
  {
   "lang": "en",
   "variant": "v3a fine-tune (unfiltered data), 8-bit",
   "shipped": false,
   "reference_all": 65,
   "variant_all": 55,
   "delta": -10,
   "reference_fabrications": 8,
   "variant_fabrications": 20,
   "scores": "diag/scores-diag-v3a-q8.json"
  },
  {
   "lang": "en",
   "variant": "v3c fine-tune (concise rewritten answers), 8-bit",
   "shipped": false,
   "reference_all": 63,
   "variant_all": 63,
   "delta": 0,
   "reference_fabrications": 6,
   "variant_fabrications": 6,
   "scores": "diag/scores-en-v3c-q8.json"
  },
  {
   "lang": "tr",
   "variant": "v3c fine-tune (concise rewritten answers), 8-bit",
   "shipped": false,
   "reference_all": 12,
   "variant_all": 37,
   "delta": 25,
   "reference_fabrications": 19,
   "variant_fabrications": 17,
   "scores": "diag/scores-tr-v3c-q8.json"
  }
 ]
}