training_config.json

{
 "model": "Sieve-9B-Plus",
 "how_it_was_made": "One LoRA run with a pointer head trained from scratch on data/train.jsonl; the checkpoint was chosen on the calibration split and the per-category temperatures fitted there.",
 "training_data": {
  "published": false,
  "files": {
   "train.jsonl": {
    "sha256": "45521bf5d9e29dd2d2b823c67d2ee8cd97e458a52847404ca245e3329c6ff8f8",
    "records": 165832
   },
   "calibration.jsonl": {
    "sha256": "e8dfe0fe0fb3cb11b2b2fa5499f32ae84a63a43c58fc0bb146ade98decba6c9c",
    "records": 30722
   },
   "test.jsonl": {
    "sha256": "aa199a3a44a26ff725a84aa0ad13e4e47288e1a8e4cec016d44f65ed07c602ca",
    "records": 116860
   }
  }
 },
 "recipe": {
  "base_model": "Qwen/Qwen3.5-9B",
  "base_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
  "adapter": {
   "type": "LoRA",
   "rank": 32,
   "alpha": 64,
   "dropout": 0.05,
   "target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
    "in_proj_qkv",
    "in_proj_z",
    "in_proj_a",
    "in_proj_b",
    "out_proj"
   ]
  },
  "pointer_head": {
   "projection_dim": 256,
   "initialisation": "from scratch",
   "dtype": "float32"
  },
  "optimiser": {
   "name": "AdamW",
   "learning_rate": 5e-05,
   "pointer_head_learning_rate": 5e-05,
   "beta2": 0.999,
   "beta1": "cycled 0.85-0.95 by the OneCycle schedule",
   "weight_decay": 0.01,
   "gradient_clip_norm": 1.0
  },
  "schedule": {
   "name": "OneCycle",
   "warmup_fraction": 0.1,
   "beta1_cycled": [
    0.85,
    0.95
   ]
  },
  "epochs": 1,
  "records_per_step": 32,
  "micro_batch_token_budget": 16384,
  "precision": "bf16 backbone, fp32 LoRA and pointer head, gradient checkpointing",
  "max_state_tokens": 16384,
  "max_question_and_options_tokens": 16384,
  "over_long_records": "dropped (never truncated)",
  "loss": "cross-entropy over each question's options (soft-target cross-entropy where a record gives a target distribution), per-question weights; a record's questions are averaged and every record weighs the same, its augmented variants sharing its weight",
  "augmentation": {
   "choice_options_permuted": true,
   "yes_no_and_rating_scale_options": "kept in their served order",
   "p_none_of_the_above_as_answer": 0.1,
   "p_none_of_the_above_as_wrong_option": 0.12,
   "p_irrelevant_distractor_option": 0.15,
   "p_none_of_the_above_minimal_pair": 0.25,
   "not_augmented": "soft targets, and questions that already offer a none/other option"
  },
  "random_seed": 0
 },
 "selection": "highest accuracy on a fixed 2000-question sample of the calibration split, lower NLL breaks ties; fixed before training",
 "temperature": {
  "value": 1.1089999999999935,
  "fit": "negative log-likelihood on the calibration split (43176 questions of the training mix, disjoint from train and test); that split is used only to fit the temperatures and choose the checkpoint; 1.0 = raw logits; the global T is the fallback for a question category without its own temperature",
  "per_category": {
   "rule": "type_options_state",
   "table": {
    "choice": 1.1089999999999935,
    "choice:10-25": 1.2439999999999953,
    "choice:10-25:empty": 1.0839999999999963,
    "choice:10-25:struct": 1.3379999999999959,
    "choice:10-25:text": 1.1509999999999945,
    "choice:2": 1.1439999999999952,
    "choice:26+": 1.1319999999999966,
    "choice:26+:empty": 1.1519999999999944,
    "choice:26+:text": 1.0979999999999948,
    "choice:2:empty": 1.1299999999999968,
    "choice:2:struct": 1.1029999999999942,
    "choice:2:text": 1.214999999999993,
    "choice:3-4": 1.0829999999999964,
    "choice:3-4:empty": 1.1209999999999922,
    "choice:3-4:struct": 1.1309999999999967,
    "choice:3-4:text": 1.0,
    "choice:5-9": 1.0489999999999946,
    "choice:5-9:struct": 1.026999999999997,
    "choice:5-9:text": 1.0599999999999934,
    "noul": 1.1089999999999935,
    "noul:2": 1.1089999999999935,
    "noul:2:struct": 1.0839999999999963,
    "noul:2:text": 1.128999999999997
   },
   "questions": {
    "choice": 34192,
    "choice:10-25": 3662,
    "choice:10-25:empty": 373,
    "choice:10-25:struct": 1297,
    "choice:10-25:text": 1992,
    "choice:2": 8947,
    "choice:26+": 1376,
    "choice:26+:empty": 572,
    "choice:26+:text": 644,
    "choice:2:empty": 3856,
    "choice:2:struct": 2975,
    "choice:2:text": 2116,
    "choice:3-4": 13178,
    "choice:3-4:empty": 4803,
    "choice:3-4:struct": 4165,
    "choice:3-4:text": 4210,
    "choice:5-9": 7029,
    "choice:5-9:struct": 2751,
    "choice:5-9:text": 4278,
    "noul": 8984,
    "noul:2": 8984,
    "noul:2:struct": 4396,
    "noul:2:text": 4588
   },
   "min_questions": 200,
   "fit": "negative log-likelihood per category on the calibration split; finest fitted category first, else the global T",
   "cross_check": {
    "global": {
     "nll": 0.30776092498926394,
     "ece": 0.008417002161563051
    },
    "per_category": {
     "nll": 0.3078539597175268,
     "ece": 0.008695022534821701
    }
   }
  }
 },
 "selected_checkpoint": {
  "step": 4500,
  "eval": {
   "step": 4500,
   "acc": 0.8835,
   "nll": 0.3066649654857755,
   "n": 2000
  }
 }
}